armor-agent-plugin 6.0.0__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Potentially problematic release.


This version of armor-agent-plugin might be problematic. Click here for more details.

Files changed (50) hide show
  1. armor/__init__.py +9 -0
  2. armor/__main__.py +3 -0
  3. armor/cli.py +209 -0
  4. armor/data/ai_standards/_catalog/latest.json +7 -0
  5. armor/data/ai_standards/_catalog/versions/2026-05-14__b1ed6db3/guardrail_catalog.json +1517 -0
  6. armor/data/ai_standards/_catalog/versions/2026-05-14__b1ed6db3/guardrail_catalog.md +345 -0
  7. armor/data/ai_standards/_catalog/versions/2026-05-14__b1ed6db3/guardrail_matrix.csv +54 -0
  8. armor/data/ai_standards/_catalog/versions/2026-05-14__b1ed6db3/guardrail_matrix.json +1446 -0
  9. armor/data/ai_standards/_catalog/versions/2026-05-14__b1ed6db3/guardrail_matrix.md +57 -0
  10. armor/data/ai_standards/_catalog/versions/2026-05-14__b1ed6db3/manifest.json +29 -0
  11. armor/data/ai_standards/_catalog/versions/2026-05-14__b1ed6db3/source_coverage.json +360 -0
  12. armor/data/ai_standards/eu-ai-act/history.jsonl +1 -0
  13. armor/data/ai_standards/eu-ai-act/latest.json +7 -0
  14. armor/data/ai_standards/eu-ai-act/versions/2026-06-04__e1a1c700/manifest.json +5 -0
  15. armor/data/ai_standards/mitre-atlas/history.jsonl +1 -0
  16. armor/data/ai_standards/mitre-atlas/latest.json +7 -0
  17. armor/data/ai_standards/mitre-atlas/versions/2026-06-04__32346fb3/manifest.json +5 -0
  18. armor/data/ai_standards/nist-ai-rmf/history.jsonl +1 -0
  19. armor/data/ai_standards/nist-ai-rmf/latest.json +7 -0
  20. armor/data/ai_standards/nist-ai-rmf/versions/2026-06-04__b157a100/manifest.json +5 -0
  21. armor/data/ai_standards/owasp-llm-top10/history.jsonl +1 -0
  22. armor/data/ai_standards/owasp-llm-top10/latest.json +7 -0
  23. armor/data/ai_standards/owasp-llm-top10/versions/2026-06-04__0a592025/manifest.json +5 -0
  24. armor/file_lock.py +114 -0
  25. armor/pipeline/__init__.py +17 -0
  26. armor/pipeline/adversary.py +140 -0
  27. armor/pipeline/conform.py +259 -0
  28. armor/pipeline/conform_audit.py +192 -0
  29. armor/pipeline/load.py +94 -0
  30. armor/pipeline/report.py +357 -0
  31. armor/pipeline/schema.py +113 -0
  32. armor/pipeline/schemas/posture_report.schema.json +400 -0
  33. armor/pipeline/state.py +182 -0
  34. armor/pipeline/triage.py +225 -0
  35. armor/pipeline/triage_audit.py +153 -0
  36. armor_agent_plugin-6.0.0.dist-info/METADATA +574 -0
  37. armor_agent_plugin-6.0.0.dist-info/RECORD +50 -0
  38. armor_agent_plugin-6.0.0.dist-info/WHEEL +5 -0
  39. armor_agent_plugin-6.0.0.dist-info/entry_points.txt +2 -0
  40. armor_agent_plugin-6.0.0.dist-info/licenses/LICENSE +21 -0
  41. armor_agent_plugin-6.0.0.dist-info/top_level.txt +2 -0
  42. catalog/__init__.py +1 -0
  43. catalog/ai_standards_versioning.py +114 -0
  44. catalog/download_aiid_slim.py +90 -0
  45. catalog/download_eu_ai_act_cellar.py +86 -0
  46. catalog/download_mitre_atlas.py +78 -0
  47. catalog/download_nist_ai_rmf.py +77 -0
  48. catalog/download_owasp_llm_top10.py +71 -0
  49. catalog/generate_guardrail_catalog.py +990 -0
  50. catalog/refresh_catalog.py +77 -0
@@ -0,0 +1,345 @@
1
+ # ARMOR AI/Agentic Guardrail Catalog
2
+
3
+ Total families: **12** | Total sub-guardrails: **53**
4
+
5
+ ## GR-01 — Risk & impact assessment (5)
6
+
7
+ ### GR-01.1: Classify AI system risk tier and intended purpose prior to deployment, documenting reasonably foreseeable misuse and affected populations.
8
+ - **Applicability**: any system; pre-deployment
9
+ - **Verification test**: Verify that a structured risk classification configuration or manifest exists and is validated before model or agent execution.
10
+ - **Implementation hints**: Maintain a machine-readable system risk card or manifest in repo config specifying risk tier, intended purpose, and out-of-scope uses.
11
+ - **Standards**: EU AI Act Art 6, 9, 27 | NIST GOVERN-1, MAP-1, MAP-3, MAP-5 | ATLAS AML.M0017 | MIT 7.3. Lack of capability or robustness; 5.1. Overreliance and unsafe use
12
+
13
+ ### GR-01.2: Maintain a continuous risk register and iterative risk management lifecycle across design, development, and post-deployment changes.
14
+ - **Applicability**: any AI system
15
+ - **Verification test**: Verify automated validation of risk register entries or policy checks tied to model/agent configuration updates.
16
+ - **Implementation hints**: Enforce risk register schema checks in CI or policy-as-code gates when prompts, models, or tools change.
17
+ - **Standards**: EU AI Act Art 9 | NIST GOVERN-1, MANAGE-1, MANAGE-2 | MIT 7.1. AI pursuing its own goals in conflict with human goals or values
18
+
19
+ ### GR-01.3: Perform fundamental rights, safety, and downstream impact assessments for high-impact or customer-facing AI deployments.
20
+ - **Applicability**: high-risk or customer-facing AI systems
21
+ - **Verification test**: Check for impact assessment gate logic or pre-flight validation prior to enabling high-impact workflows.
22
+ - **Implementation hints**: Include impact assessment assertions and threshold gates in deployment readiness checks.
23
+ - **Standards**: EU AI Act Art 9, 27 | NIST MAP-2, MAP-5, MEASURE-2 | MIT 6.1. Power centralization and unfair distribution of benefits
24
+
25
+ ### GR-01.4: Define explicit operational boundaries, capability limits, and refusal conditions for out-of-domain requests.
26
+ - **Applicability**: all LLM and generative AI systems
27
+ - **Verification test**: Verify input routing or domain boundary enforcement that rejects or deflects out-of-scope queries.
28
+ - **Implementation hints**: Implement domain/topic classifiers or boundary guards in the request intake pipeline before invoking the model.
29
+ - **Standards**: EU AI Act Art 9, 13 | NIST MAP-3, MANAGE-3 | OWASP LLM01, LLM09 | ATLAS AML.M0015 | MIT 5.1. Overreliance and unsafe use
30
+
31
+ ### GR-01.5: Gate release and runtime activation on quantitative risk acceptance thresholds and sign-off controls.
32
+ - **Applicability**: any production AI system
33
+ - **Verification test**: Verify threshold-check functions that block execution or deployment when evaluation metrics fall below required bounds.
34
+ - **Implementation hints**: Enforce automated evaluation score thresholds before promoting models, prompts, or agent workflows.
35
+ - **Standards**: EU AI Act Art 9, 43 | NIST GOVERN-4, MEASURE-4, MANAGE-4 | MIT 7.3. Lack of capability or robustness
36
+
37
+ ## GR-02 — Data governance (5)
38
+
39
+ ### GR-02.1: Track dataset, corpus, and retrieval index provenance, lineage, and licensing metadata across ingestion pipelines.
40
+ - **Applicability**: systems ingesting training, fine-tuning, or RAG data
41
+ - **Verification test**: Verify that ingestion and indexing functions record cryptographic hashes and provenance metadata per document.
42
+ - **Implementation hints**: Attach source URI, content SHA256 hash, timestamp, and license fields to ingested documents and vector store records.
43
+ - **Standards**: EU AI Act Art 10, 53 | NIST MAP-2, MEASURE-2 | OWASP LLM04, LLM08 | ATLAS AML.M0007 | MIT 4.3. Lack of transparency or interpretability
44
+
45
+ ### GR-02.2: Validate data quality, schema conformance, and completeness prior to indexing, embedding, or model consumption.
46
+ - **Applicability**: systems ingesting external or structured data
47
+ - **Verification test**: Confirm presence of schema validation and rejection handling in data loading and RAG ingestion paths.
48
+ - **Implementation hints**: Use strict schema validation (e.g., Pydantic) and null/corruption checks on all ingested records.
49
+ - **Standards**: EU AI Act Art 10 | NIST MEASURE-2, MANAGE-2 | OWASP LLM04, LLM08 | ATLAS AML.M0007, AML.M0014 | MIT 7.3. Lack of capability or robustness
50
+
51
+ ### GR-02.3: Scan training, fine-tuning, and RAG corpora for data poisoning, backdoor triggers, and anomalous payloads.
52
+ - **Applicability**: RAG, fine-tuning, and knowledge-grounded systems
53
+ - **Verification test**: Verify sanitization and poisoning detection filters execute on untrusted corpus inputs prior to storage.
54
+ - **Implementation hints**: Run anomaly detection, sanitization filters, and integrity verification over documents before embedding.
55
+ - **Standards**: EU AI Act Art 10, 15 | NIST MEASURE-2, MANAGE-3 | OWASP LLM04, LLM08 | ATLAS AML.M0006, AML.M0007, AML.M0014 | MIT 2.2. AI system security vulnerabilities and attacks
56
+
57
+ ### GR-02.4: Enforce tenant isolation and access-control filtering on vector stores, embeddings, and retrieved context chunks.
58
+ - **Applicability**: multi-tenant or access-controlled RAG systems
59
+ - **Verification test**: Verify that vector/document retrieval functions require and filter by caller authorization scope or tenant ID.
60
+ - **Implementation hints**: Bind caller principal/ACL attributes to retrieval queries so cross-tenant chunks are never injected into prompts.
61
+ - **Standards**: EU AI Act Art 10 | NIST GOVERN-6, MANAGE-3 | OWASP LLM02, LLM08 | ATLAS AML.M0018 | MIT 1.2. Privacy violations
62
+
63
+ ### GR-02.5: Implement retention schedules, right-to-erasure purge routines, and versioned snapshots for datasets and embeddings.
64
+ - **Applicability**: systems persisting user or enterprise data
65
+ - **Verification test**: Verify existence of TTL/retention enforcement or document deletion routines in storage modules.
66
+ - **Implementation hints**: Provide deterministic delete-by-source-ID and TTL expiration jobs across document stores and caches.
67
+ - **Standards**: EU AI Act Art 10 | NIST GOVERN-1, MANAGE-4 | OWASP LLM02 | MIT 1.2. Privacy violations
68
+
69
+ ## GR-03 — Transparency & disclosure (4)
70
+
71
+ ### GR-03.1: Disclose clearly to users and downstream consumers when they are interacting with an AI system or reading AI-generated content.
72
+ - **Applicability**: user-facing AI and content generation systems
73
+ - **Verification test**: Verify response envelopes or UI rendering templates include an explicit AI-generated disclosure indicator.
74
+ - **Implementation hints**: Attach explicit AI disclosure headers, metadata flags, or UI notices to all generated responses.
75
+ - **Standards**: EU AI Act Art 13, 50 | NIST GOVERN-4, MAP-1, MEASURE-2 | OWASP LLM09 | MIT 4.3. Lack of transparency or interpretability; 3.1. Pollution of information ecosystem and loss of shared reality
76
+
77
+ ### GR-03.2: Ground generated claims in verifiable source citations and cross-check citations against the underlying corpus or files.
78
+ - **Applicability**: RAG, coding, compliance, and analytical AI systems
79
+ - **Verification test**: Verify post-generation citation validation logic that checks cited sources exist and demotes or flags unverified claims.
80
+ - **Implementation hints**: Require structured citations on model outputs and deterministically verify file/document existence and line/span bounds.
81
+ - **Standards**: EU AI Act Art 13 | NIST MEASURE-2, MANAGE-2 | OWASP LLM09 | MIT 3.2. False or misleading information; 4.3. Lack of transparency or interpretability
82
+
83
+ ### GR-03.3: Provide model/system cards documenting model architecture, version, known limitations, and evaluation benchmarks.
84
+ - **Applicability**: any AI system
85
+ - **Verification test**: Verify presence of structured system metadata, version identifiers, and documented limitations.
86
+ - **Implementation hints**: Ship structured system/model metadata alongside code and expose version/capability descriptors via CLI or API.
87
+ - **Standards**: EU AI Act Art 11, 13, 53 | NIST GOVERN-1, MAP-1, MEASURE-4 | MIT 4.3. Lack of transparency or interpretability
88
+
89
+ ### GR-03.4: Expose interpretable reasoning traces, confidence indicators, or uncertainty warnings on automated assessments.
90
+ - **Applicability**: decision-support and analytical AI systems
91
+ - **Verification test**: Verify output schemas require explicit reasoning and support inconclusive/uncertainty states rather than forced binary verdicts.
92
+ - **Implementation hints**: Include structured reasoning fields, uncertainty/inconclusive states, and audit trails in assessment outputs.
93
+ - **Standards**: EU AI Act Art 13, 14 | NIST MEASURE-2, MANAGE-1 | OWASP LLM09 | MIT 4.3. Lack of transparency or interpretability; 5.1. Overreliance and unsafe use
94
+
95
+ ## GR-04 — Human oversight (4)
96
+
97
+ ### GR-04.1: Require human-in-the-loop (HITL) approval before executing high-impact, irreversible, or state-mutating actions.
98
+ - **Applicability**: agentic and tool-using AI systems
99
+ - **Verification test**: Verify that state-mutating or high-risk actions check for explicit user/operator approval before execution.
100
+ - **Implementation hints**: Gate destructive, financial, or external write tool calls behind an explicit confirmation token or approval queue.
101
+ - **Standards**: EU AI Act Art 14 | NIST GOVERN-3, MAP-3, MANAGE-2 | OWASP LLM06 | ATLAS AML.M0018 | MIT 5.2. Loss of human agency and autonomy; 7.2. AI possessing dangerous capabilities
102
+
103
+ ### GR-04.2: Provide an immediate stop, cancellation, or kill-switch mechanism to halt autonomous loops and background tasks.
104
+ - **Applicability**: autonomous and multi-step AI pipelines
105
+ - **Verification test**: Verify bounded loop counters (e.g., max revisions/steps) and clean abort handling in execution drivers.
106
+ - **Implementation hints**: Enforce hard iteration caps, timeouts, and signal/cancellation checks inside agent and audit loops.
107
+ - **Standards**: EU AI Act Art 14 | NIST GOVERN-3, MANAGE-3 | OWASP LLM06, LLM10 | MIT 7.4. Lack of capability to control AI; 5.2. Loss of human agency and autonomy
108
+
109
+ ### GR-04.3: Support human override, correction, and escalation queues for contested or ambiguous model verdicts.
110
+ - **Applicability**: decision-support and auditing AI systems
111
+ - **Verification test**: Verify that unconverged or contested items are collected into an explicit open_questions or escalation structure.
112
+ - **Implementation hints**: Surface unresolved disputes as explicit open_questions for human review rather than silently masking disagreements.
113
+ - **Standards**: EU AI Act Art 14 | NIST GOVERN-3, MANAGE-1, MANAGE-4 | OWASP LLM09 | MIT 5.1. Overreliance and unsafe use
114
+
115
+ ### GR-04.4: Prevent automation bias by presenting counter-evidence, adversarial critiques, and verification status alongside conclusions.
116
+ - **Applicability**: analytical, compliance, and recommendation AI systems
117
+ - **Verification test**: Verify presence of adversarial review/critique stages and explicit demotion markers in user-facing reports.
118
+ - **Implementation hints**: Run an independent adversarial critique pass and highlight demoted or weakly supported claims in reports.
119
+ - **Standards**: EU AI Act Art 14 | NIST MAP-3, MEASURE-2 | OWASP LLM09 | MIT 5.1. Overreliance and unsafe use
120
+
121
+ ## GR-05 — Robustness & accuracy (5)
122
+
123
+ ### GR-05.1: Validate all LLM outputs against strict structural and type schemas before downstream consumption or persistence.
124
+ - **Applicability**: all structured-output LLM and agentic systems
125
+ - **Verification test**: Verify strict schema validation on LLM responses and final report artifacts prior to writing or acting on them.
126
+ - **Implementation hints**: Parse and validate every LLM response against JSON Schema or Pydantic models with extra='forbid'; reject malformed payloads.
127
+ - **Standards**: EU AI Act Art 15 | NIST MEASURE-2, MANAGE-2 | OWASP LLM05, LLM09 | ATLAS AML.M0015 | MIT 7.3. Lack of capability or robustness
128
+
129
+ ### GR-05.2: Detect and mitigate hallucinations by cross-verifying model assertions against deterministic ground-truth sources.
130
+ - **Applicability**: all generative AI systems making factual or code assertions
131
+ - **Verification test**: Verify deterministic cross-verification of model citations/facts and automatic demotion + logging on failure.
132
+ - **Implementation hints**: Resolve file paths, line numbers, or database keys cited by the LLM and automatically demote unverified claims.
133
+ - **Standards**: EU AI Act Art 15 | NIST MEASURE-2, MANAGE-2 | OWASP LLM09 | MIT 3.2. False or misleading information; 7.3. Lack of capability or robustness
134
+
135
+ ### GR-05.3: Handle malformed inputs, transient failures, and partial responses gracefully with deterministic error codes and recovery paths.
136
+ - **Applicability**: all AI systems
137
+ - **Verification test**: Verify deterministic validation error handling and non-zero exit codes when inputs or responses are missing or invalid.
138
+ - **Implementation hints**: Wrap file/JSON parsing and phase transitions in explicit error handlers that emit actionable error messages and non-zero exit codes.
139
+ - **Standards**: EU AI Act Art 15 | NIST MEASURE-2, MANAGE-3 | OWASP LLM05, LLM10 | MIT 7.3. Lack of capability or robustness
140
+
141
+ ### GR-05.4: Maintain an automated regression and unit test suite covering pipeline invariants, edge cases, and failure modes.
142
+ - **Applicability**: all AI systems
143
+ - **Verification test**: Verify existence of automated tests covering schema validation, citation verification, and pipeline phase transitions.
144
+ - **Implementation hints**: Ship automated pytest tests verifying happy paths, invalid inputs, hallucination demotions, and schema drift.
145
+ - **Standards**: EU AI Act Art 15, 43 | NIST MEASURE-1, MEASURE-2, MEASURE-4 | MIT 7.3. Lack of capability or robustness
146
+
147
+ ### GR-05.5: Enforce multi-agent or adversarial cross-checking to reduce single-pass model blind spots and drift.
148
+ - **Applicability**: high-assurance analytical and agentic systems
149
+ - **Verification test**: Verify implementation of adversarial audit phases that critique and apply deltas to primary LLM outputs.
150
+ - **Implementation hints**: Pair primary assessment phases with bounded adversarial red-team audit phases that can CONFIRM, REVISE, or INVESTIGATE.
151
+ - **Standards**: EU AI Act Art 15 | NIST MEASURE-2, MANAGE-2 | OWASP LLM09 | ATLAS AML.M0003 | MIT 7.3. Lack of capability or robustness
152
+
153
+ ## GR-06 — Adversarial security (5)
154
+
155
+ ### GR-06.1: Defend against direct and indirect prompt injection by separating untrusted data from system instructions and constraining model outputs.
156
+ - **Applicability**: all LLM and agentic systems processing untrusted inputs
157
+ - **Verification test**: Verify that untrusted inputs are structured separately from instructions and responses are strictly validated as data.
158
+ - **Implementation hints**: Isolate untrusted repository content into structured data fields, forbid prose execution, and require strict JSON schema adherence.
159
+ - **Standards**: EU AI Act Art 15 | NIST MAP-3, MEASURE-2, MANAGE-3 | OWASP LLM01, LLM08 | ATLAS AML.M0015, AML.M0019 | MIT 2.2. AI system security vulnerabilities and attacks
160
+
161
+ ### GR-06.2: Sanitize and validate all model-generated strings, paths, and parameters before passing them to OS, filesystem, shell, or database APIs.
162
+ - **Applicability**: all AI systems interacting with filesystems, shells, or APIs
163
+ - **Verification test**: Verify path traversal checks (relative_to project_root) and absence of unsanitized shell interpolation.
164
+ - **Implementation hints**: Resolve paths against project_root, reject '..' traversal, avoid shell=True, and validate identifiers against strict regexes.
165
+ - **Standards**: EU AI Act Art 15 | NIST MEASURE-2, MANAGE-3 | OWASP LLM05, LLM06 | ATLAS AML.M0015, AML.M0018 | MIT 2.2. AI system security vulnerabilities and attacks
166
+
167
+ ### GR-06.3: Enforce resource bounds, input size limits, and loop caps to prevent denial-of-service or unbounded token consumption.
168
+ - **Applicability**: all LLM and agentic systems
169
+ - **Verification test**: Verify explicit character/file limits on context gathering and hard upper bounds on retry/revision loops.
170
+ - **Implementation hints**: Truncate ingested manifest/doc previews (e.g., 6000 chars), cap file walks (e.g., 200 files), and bound audit revisions.
171
+ - **Standards**: EU AI Act Art 15 | NIST MANAGE-3 | OWASP LLM10 | ATLAS AML.M0004 | MIT 2.2. AI system security vulnerabilities and attacks
172
+
173
+ ### GR-06.4: Prevent system prompt leakage and sensitive internal context exposure in user-facing or downstream outputs.
174
+ - **Applicability**: all LLM systems
175
+ - **Verification test**: Verify that external output contracts exclude raw system prompts, credentials, and unrestricted source dumps.
176
+ - **Implementation hints**: Emit only validated contract fields in final reports and agent wrappers; strip internal prompts and raw source excerpts.
177
+ - **Standards**: EU AI Act Art 15 | NIST MEASURE-2, MANAGE-3 | OWASP LLM02, LLM07 | ATLAS AML.M0002, AML.M0015 | MIT 2.2. AI system security vulnerabilities and attacks; 1.2. Privacy violations
178
+
179
+ ### GR-06.5: Conduct regular adversarial red-teaming against model verdicts and evidence claims under strict read-scope constraints.
180
+ - **Applicability**: all high-assurance AI systems
181
+ - **Verification test**: Verify adversary instructions and delta application forbid inventing new IDs or expanding scope.
182
+ - **Implementation hints**: Run dedicated adversary phases constrained to auditing existing citations without inventing new scope.
183
+ - **Standards**: EU AI Act Art 15, 55 | NIST MEASURE-2, MANAGE-2 | OWASP LLM01, LLM09 | ATLAS AML.M0003 | MIT 2.2. AI system security vulnerabilities and attacks
184
+
185
+ ## GR-07 — Privacy & data protection (4)
186
+
187
+ ### GR-07.1: Minimize sensitive and personal data exposure by restricting context collection to necessary manifests and bounded file excerpts.
188
+ - **Applicability**: all AI systems accessing user or repository data
189
+ - **Verification test**: Verify that context collection skips dotfiles/environments and enforces strict character and file count limits.
190
+ - **Implementation hints**: Read only curated manifest filenames with character caps and skip hidden/credential directories during project walks.
191
+ - **Standards**: EU AI Act Art 10, 15 | NIST GOVERN-1, MAP-2, MANAGE-3 | OWASP LLM02 | ATLAS AML.M0012 | MIT 1.2. Privacy violations
192
+
193
+ ### GR-07.2: Prevent unauthorized outbound network exfiltration of codebase or user data during inference and pipeline execution.
194
+ - **Applicability**: all enterprise and local-first AI tools
195
+ - **Verification test**: Verify that runtime pipeline modules make zero network requests and never import or invoke external HTTP/LLM clients.
196
+ - **Implementation hints**: Execute assessment pipelines purely locally via disk-queue exchanges without opening outbound network sockets.
197
+ - **Standards**: EU AI Act Art 15 | NIST GOVERN-6, MANAGE-3 | OWASP LLM02, LLM06 | ATLAS AML.M0018 | MIT 1.2. Privacy violations; 2.2. AI system security vulnerabilities and attacks
198
+
199
+ ### GR-07.3: Scrub or exclude secrets, API keys, and credentials from logs, artifacts, and model context windows.
200
+ - **Applicability**: all AI systems
201
+ - **Verification test**: Verify that .env and secret files are excluded from manifest candidate lists and no credentials appear in artifacts.
202
+ - **Implementation hints**: Never read .env or credential files into prompts and avoid hardcoding or logging tokens in pipeline state.
203
+ - **Standards**: EU AI Act Art 10, 15 | NIST MANAGE-3 | OWASP LLM02 | MIT 1.2. Privacy violations
204
+
205
+ ### GR-07.4: Confine all runtime state, temporary files, and reports to a designated, git-ignorable workspace directory.
206
+ - **Applicability**: all local and repository-auditing AI tools
207
+ - **Verification test**: Verify that all runtime file writes target <project_root>/.armor/ (or atomic temp files in the same directory).
208
+ - **Implementation hints**: Write all run artifacts, queues, logs, and reports exclusively beneath <project_root>/.armor/.
209
+ - **Standards**: EU AI Act Art 10, 12 | NIST GOVERN-1, MANAGE-3 | OWASP LLM02 | MIT 1.2. Privacy violations
210
+
211
+ ## GR-08 — Fairness & non-discrimination (4)
212
+
213
+ ### GR-08.1: Evaluate and document affected user populations and potential disparate impacts during system profiling.
214
+ - **Applicability**: all AI systems
215
+ - **Verification test**: Verify that the triage system_profile schema requires explicit user_population profiling.
216
+ - **Implementation hints**: Require user_population and data_sensitivity fields in the Phase 2 system_profile schema.
217
+ - **Standards**: EU AI Act Art 10, 13, 27 | NIST MAP-1, MAP-2, MEASURE-2 | MIT 1.1. Unfair discrimination and misrepresentation
218
+
219
+ ### GR-08.2: Apply consistent, deterministic scoring weights and evaluation criteria across all assessed components or subjects.
220
+ - **Applicability**: scoring, ranking, and auditing AI systems
221
+ - **Verification test**: Verify that coverage and suggestion rankings are computed deterministically from a fixed weight map.
222
+ - **Implementation hints**: Compute posture scores using a fixed, transparent weight table (VERDICT_WEIGHT) and deterministic sorting.
223
+ - **Standards**: EU AI Act Art 10, 15 | NIST MEASURE-1, MEASURE-2 | MIT 1.1. Unfair discrimination and misrepresentation
224
+
225
+ ### GR-08.3: Exclude inapplicable criteria from scoring denominators so specialized systems are not unfairly penalized.
226
+ - **Applicability**: evaluation and compliance scoring systems
227
+ - **Verification test**: Verify that 'n/a' verdicts are excluded from the coverage denominator in report rollup calculations.
228
+ - **Implementation hints**: Exclude rows with verdict == 'n/a' from in_scope_count and max_possible_points when computing coverage percentages.
229
+ - **Standards**: EU AI Act Art 9, 15 | NIST MEASURE-1, MEASURE-2 | MIT 1.1. Unfair discrimination and misrepresentation
230
+
231
+ ### GR-08.4: Audit subjective classifications through an independent review pass to detect biased or unsubstantiated scoping decisions.
232
+ - **Applicability**: classification and triage AI systems
233
+ - **Verification test**: Verify that triage applicability decisions are audited by triage-audit before conformance scoring proceeds.
234
+ - **Implementation hints**: Subject all Phase 2 applicability tags (applies/partial_scope/n_a) to the Phase 3 triage-audit red-team loop.
235
+ - **Standards**: EU AI Act Art 10, 14 | NIST MEASURE-2, MANAGE-2 | MIT 1.1. Unfair discrimination and misrepresentation
236
+
237
+ ## GR-09 — Content safety (4)
238
+
239
+ ### GR-09.1: Enforce output content quality and forbid vague, boilerplate, or non-actionable safety/remediation advice.
240
+ - **Applicability**: advisory, code-review, and remediation-generating AI systems
241
+ - **Verification test**: Verify that conformance instructions mandate codebase-specific, action-leading remediations and list banned phrases.
242
+ - **Implementation hints**: Instruct the model with explicit remediation quality rules and a list of banned boilerplate phrases.
243
+ - **Standards**: EU AI Act Art 13, 15 | NIST MEASURE-2, MANAGE-2 | OWASP LLM05, LLM09 | MIT 5.1. Overreliance and unsafe use
244
+
245
+ ### GR-09.2: Escape and sanitize untrusted or model-generated strings before rendering them into Markdown tables or reports.
246
+ - **Applicability**: systems generating Markdown or HTML reports from LLM outputs
247
+ - **Verification test**: Verify Markdown rendering functions escape '|' and newlines in table cells.
248
+ - **Implementation hints**: Escape pipe characters ('|') and normalize newlines when rendering model reasoning and actions into Markdown tables.
249
+ - **Standards**: EU AI Act Art 15 | NIST MANAGE-3 | OWASP LLM05 | MIT 2.2. AI system security vulnerabilities and attacks
250
+
251
+ ### GR-09.3: Provide safe, deterministic fallback guidance whenever model-generated remediation text is missing or inconclusive.
252
+ - **Applicability**: advisory and reporting AI systems
253
+ - **Verification test**: Verify report suggestion builder provides explicit fallback action strings when remediation is empty.
254
+ - **Implementation hints**: Supply deterministic fallback action strings for inconclusive or empty-remediation rows during report generation.
255
+ - **Standards**: EU AI Act Art 13, 15 | NIST MANAGE-2, MANAGE-3 | OWASP LLM09 | MIT 5.1. Overreliance and unsafe use
256
+
257
+ ### GR-09.4: Constrain agent wrapper summaries to concise, bounded notes free of raw source code, citations, or prompt instructions.
258
+ - **Applicability**: agent wrappers and automated CI/CD gates
259
+ - **Verification test**: Verify agent wrapper specification enforces bounded notes (<200 chars) with no source excerpts or prompts.
260
+ - **Implementation hints**: Enforce a strict <200 character single-line summary note in the agent YAML verdict contract.
261
+ - **Standards**: EU AI Act Art 13, 15 | NIST GOVERN-4, MANAGE-3 | OWASP LLM02, LLM05 | MIT 1.2. Privacy violations
262
+
263
+ ## GR-10 — Logging, monitoring & incident response (5)
264
+
265
+ ### GR-10.1: Maintain an immutable, timestamped run directory recording the input/output artifact of every pipeline phase.
266
+ - **Applicability**: all production and auditing AI pipelines
267
+ - **Verification test**: Verify that new_run creates UTC-timestamped run directories and each phase persists its numbered artifact.
268
+ - **Implementation hints**: Create UTC-timestamped directories under .armor/runs/<iso>/ and persist numbered JSON/MD artifacts for Phases 1–6.
269
+ - **Standards**: EU AI Act Art 12, 19 | NIST GOVERN-1, MEASURE-2, MANAGE-4 | ATLAS AML.M0016 | MIT 4.3. Lack of transparency or interpretability
270
+
271
+ ### GR-10.2: Log every detected model hallucination or rejected citation to an append-only structured incident log.
272
+ - **Applicability**: all citation-backed generative AI systems
273
+ - **Verification test**: Verify that rejected citations in conform and conform-audit are appended to hallucination_log.jsonl.
274
+ - **Implementation hints**: Append JSONL records with UTC timestamp, phase, sid, citation, and rejection reason to .armor/hallucination_log.jsonl.
275
+ - **Standards**: EU AI Act Art 12, 73 | NIST MEASURE-2, MANAGE-4 | OWASP LLM09 | ATLAS AML.M0016 | MIT 3.2. False or misleading information; 4.3. Lack of transparency or interpretability
276
+
277
+ ### GR-10.3: Record full multi-round adversarial audit trails capturing per-round verdict counts, critiques, and before/after row diffs.
278
+ - **Applicability**: systems with multi-step or adversarial review loops
279
+ - **Verification test**: Verify that audit phases record per-round verdict_counts and before/after diffs in audit_trail.
280
+ - **Implementation hints**: Persist audit_trail arrays in 03_triage_audit.json and 05_conformance_audit.json with before/after revision records.
281
+ - **Standards**: EU AI Act Art 12 | NIST MEASURE-2, MANAGE-4 | MIT 4.3. Lack of transparency or interpretability
282
+
283
+ ### GR-10.4: Enforce mandatory skill-usage telemetry tracking across parent skills and delegated sub-agents.
284
+ - **Applicability**: marketplace skills and multi-agent plugins
285
+ - **Verification test**: Verify that SKILL.md defines a mandatory telemetry gate before tool calls, sub-agents, and final responses.
286
+ - **Implementation hints**: Gate skill execution on emitting start and end telemetry events with propagated PARENT_TRACKERID.
287
+ - **Standards**: EU AI Act Art 12 | NIST GOVERN-1, MANAGE-4 | ATLAS AML.M0016 | MIT 4.3. Lack of transparency or interpretability
288
+
289
+ ### GR-10.5: Expose a deterministic CLI status inspection command reporting live progress across all pipeline phases.
290
+ - **Applicability**: multi-phase AI pipelines
291
+ - **Verification test**: Verify that `pipeline <project> --status` reports the current run root and the state of every phase.
292
+ - **Implementation hints**: Support `--status` on the pipeline CLI to inspect whether each phase is pending, awaiting-llm, response-ready, or done.
293
+ - **Standards**: EU AI Act Art 12, 14 | NIST GOVERN-3, MANAGE-4 | MIT 4.3. Lack of transparency or interpretability
294
+
295
+ ## GR-11 — Supply-chain & component integrity (4)
296
+
297
+ ### GR-11.1: Version and content-address all external standards, catalogs, and reference artifacts using cryptographic SHA256 hashes.
298
+ - **Applicability**: systems depending on external catalogs, standards, or model artifacts
299
+ - **Verification test**: Verify content-addressed SHA256 hashing and manifest/history tracking in ai_standards_versioning.py.
300
+ - **Implementation hints**: Commit catalog and standard updates via content-addressed `<date>__<sha8>` directories, `latest.json`, and `history.jsonl`.
301
+ - **Standards**: EU AI Act Art 15, 53 | NIST GOVERN-6, MAP-2, MANAGE-3 | OWASP LLM03 | ATLAS AML.M0007, AML.M0014 | MIT 2.2. AI system security vulnerabilities and attacks
302
+
303
+ ### GR-11.2: Compute deterministic source fingerprints of the target codebase to detect concurrent edits or environment drift.
304
+ - **Applicability**: codebase analysis and auditing tools
305
+ - **Verification test**: Verify that project_fingerprint() produces a deterministic SHA256 prefix that changes whenever a source file is edited.
306
+ - **Implementation hints**: Hash sorted relative paths and modification timestamps across tracked source extensions via project_fingerprint().
307
+ - **Standards**: EU AI Act Art 15 | NIST MEASURE-2, MANAGE-3 | OWASP LLM03 | ATLAS AML.M0014 | MIT 7.3. Lack of capability or robustness
308
+
309
+ ### GR-11.3: Minimize third-party runtime dependencies and isolate network-fetching utilities from the core execution runtime.
310
+ - **Applicability**: all software and AI plugins
311
+ - **Verification test**: Verify that core `armor/` runtime modules import only stdlib and pydantic.
312
+ - **Implementation hints**: Require only pydantic at runtime (stdlib for everything else); isolate requests/pyyaml to offline catalog refresh scripts.
313
+ - **Standards**: EU AI Act Art 15 | NIST GOVERN-6, MANAGE-3 | OWASP LLM03 | ATLAS AML.M0014 | MIT 2.2. AI system security vulnerabilities and attacks
314
+
315
+ ### GR-11.4: Keep packaged JSON Schemas byte-synchronized with runtime data models and provide automated drift verification.
316
+ - **Applicability**: all schema-driven AI pipelines
317
+ - **Verification test**: Verify `--write-schema` and test assertions that `posture_report.schema.json` byte-equals `PostureReport.model_json_schema()`.
318
+ - **Implementation hints**: Ship `posture_report.schema.json` generated from `PostureReport.model_json_schema()` and enforce byte equality in tests.
319
+ - **Standards**: EU AI Act Art 15 | NIST MEASURE-2, MANAGE-3 | OWASP LLM03, LLM05 | MIT 7.3. Lack of capability or robustness
320
+
321
+ ## GR-12 — Agentic autonomy & action safety (4)
322
+
323
+ ### GR-12.1: Enforce strict phase ordering so downstream agent actions cannot execute before prerequisite artifacts are verified.
324
+ - **Applicability**: multi-step agentic workflows
325
+ - **Verification test**: Verify that calling later phases before earlier phases raises SystemExit via `require_phase_done`.
326
+ - **Implementation hints**: Guard each phase entry point with `require_phase_done(run, prerequisite_phase)` and abort with exact unblock instructions.
327
+ - **Standards**: EU AI Act Art 14, 15 | NIST GOVERN-3, MANAGE-3 | OWASP LLM06 | ATLAS AML.M0018 | MIT 7.4. Lack of capability to control AI; 5.2. Loss of human agency and autonomy
328
+
329
+ ### GR-12.2: Prevent recursive self-auditing or runaway self-targeting in autonomous agent wrappers.
330
+ - **Applicability**: autonomous auditing and delegation agents
331
+ - **Verification test**: Verify that the agent wrapper specification mandates a self-targeting refusal returning `verdict: fail` and zeroed counts.
332
+ - **Implementation hints**: Enforce a mandatory self-targeting guard in `armor-agent` that refuses paths resolving to `armor-agent-plugin`.
333
+ - **Standards**: EU AI Act Art 14, 15 | NIST GOVERN-3, MANAGE-3 | OWASP LLM06, LLM10 | ATLAS AML.M0018 | MIT 7.4. Lack of capability to control AI; 7.1. AI pursuing its own goals in conflict with human goals or values
334
+
335
+ ### GR-12.3: Protect concurrent file mutations with cross-platform exclusive file locks and atomic write-and-replace primitives.
336
+ - **Applicability**: concurrent or multi-agent local workflows
337
+ - **Verification test**: Verify cross-platform locking (`msvcrt` / `fcntl` + `threading.Lock`) and atomic `os.replace` writes in `file_lock.py`.
338
+ - **Implementation hints**: Provide `exclusive_lock`, `locked_write` (mkstemp + os.replace), and `locked_append` in `file_lock.py`.
339
+ - **Standards**: EU AI Act Art 15 | NIST MANAGE-3 | OWASP LLM06 | MIT 7.3. Lack of capability or robustness
340
+
341
+ ### GR-12.4: Constrain red-team adversary agents to read-only verification of cited evidence without expanding scope or inventing findings.
342
+ - **Applicability**: multi-agent and adversarial AI systems
343
+ - **Verification test**: Verify `apply_deltas` ignores unknown SIDs (recording them in `unknown_sids`) and preserves original row ordering.
344
+ - **Implementation hints**: Instruct adversary phases not to read files beyond cited `file:line` spans and ignore any unknown SIDs in apply_deltas.
345
+ - **Standards**: EU AI Act Art 14, 15 | NIST GOVERN-3, MEASURE-2, MANAGE-3 | OWASP LLM06 | ATLAS AML.M0018 | MIT 7.4. Lack of capability to control AI
@@ -0,0 +1,54 @@
1
+ number,category,description,implementation_hints,verification_test,applicability,standard_references
2
+ GR-01.1,Risk & impact assessment,"Classify AI system risk tier and intended purpose prior to deployment, documenting reasonably foreseeable misuse and affected populations.","Maintain a machine-readable system risk card or manifest in repo config specifying risk tier, intended purpose, and out-of-scope uses.",Verify that a structured risk classification configuration or manifest exists and is validated before model or agent execution.,any system; pre-deployment,"EU AI Act Art 6, 9, 27 | NIST GOVERN-1, MAP-1, MAP-3, MAP-5 | ATLAS AML.M0017 | MIT 7.3. Lack of capability or robustness; 5.1. Overreliance and unsafe use"
3
+ GR-01.2,Risk & impact assessment,"Maintain a continuous risk register and iterative risk management lifecycle across design, development, and post-deployment changes.","Enforce risk register schema checks in CI or policy-as-code gates when prompts, models, or tools change.",Verify automated validation of risk register entries or policy checks tied to model/agent configuration updates.,any AI system,"EU AI Act Art 9 | NIST GOVERN-1, MANAGE-1, MANAGE-2 | MIT 7.1. AI pursuing its own goals in conflict with human goals or values"
4
+ GR-01.3,Risk & impact assessment,"Perform fundamental rights, safety, and downstream impact assessments for high-impact or customer-facing AI deployments.",Include impact assessment assertions and threshold gates in deployment readiness checks.,Check for impact assessment gate logic or pre-flight validation prior to enabling high-impact workflows.,high-risk or customer-facing AI systems,"EU AI Act Art 9, 27 | NIST MAP-2, MAP-5, MEASURE-2 | MIT 6.1. Power centralization and unfair distribution of benefits"
5
+ GR-01.4,Risk & impact assessment,"Define explicit operational boundaries, capability limits, and refusal conditions for out-of-domain requests.",Implement domain/topic classifiers or boundary guards in the request intake pipeline before invoking the model.,Verify input routing or domain boundary enforcement that rejects or deflects out-of-scope queries.,all LLM and generative AI systems,"EU AI Act Art 9, 13 | NIST MAP-3, MANAGE-3 | OWASP LLM01, LLM09 | ATLAS AML.M0015 | MIT 5.1. Overreliance and unsafe use"
6
+ GR-01.5,Risk & impact assessment,Gate release and runtime activation on quantitative risk acceptance thresholds and sign-off controls.,"Enforce automated evaluation score thresholds before promoting models, prompts, or agent workflows.",Verify threshold-check functions that block execution or deployment when evaluation metrics fall below required bounds.,any production AI system,"EU AI Act Art 9, 43 | NIST GOVERN-4, MEASURE-4, MANAGE-4 | MIT 7.3. Lack of capability or robustness"
7
+ GR-02.1,Data governance,"Track dataset, corpus, and retrieval index provenance, lineage, and licensing metadata across ingestion pipelines.","Attach source URI, content SHA256 hash, timestamp, and license fields to ingested documents and vector store records.",Verify that ingestion and indexing functions record cryptographic hashes and provenance metadata per document.,"systems ingesting training, fine-tuning, or RAG data","EU AI Act Art 10, 53 | NIST MAP-2, MEASURE-2 | OWASP LLM04, LLM08 | ATLAS AML.M0007 | MIT 4.3. Lack of transparency or interpretability"
8
+ GR-02.2,Data governance,"Validate data quality, schema conformance, and completeness prior to indexing, embedding, or model consumption.","Use strict schema validation (e.g., Pydantic) and null/corruption checks on all ingested records.",Confirm presence of schema validation and rejection handling in data loading and RAG ingestion paths.,systems ingesting external or structured data,"EU AI Act Art 10 | NIST MEASURE-2, MANAGE-2 | OWASP LLM04, LLM08 | ATLAS AML.M0007, AML.M0014 | MIT 7.3. Lack of capability or robustness"
9
+ GR-02.3,Data governance,"Scan training, fine-tuning, and RAG corpora for data poisoning, backdoor triggers, and anomalous payloads.","Run anomaly detection, sanitization filters, and integrity verification over documents before embedding.",Verify sanitization and poisoning detection filters execute on untrusted corpus inputs prior to storage.,"RAG, fine-tuning, and knowledge-grounded systems","EU AI Act Art 10, 15 | NIST MEASURE-2, MANAGE-3 | OWASP LLM04, LLM08 | ATLAS AML.M0006, AML.M0007, AML.M0014 | MIT 2.2. AI system security vulnerabilities and attacks"
10
+ GR-02.4,Data governance,"Enforce tenant isolation and access-control filtering on vector stores, embeddings, and retrieved context chunks.",Bind caller principal/ACL attributes to retrieval queries so cross-tenant chunks are never injected into prompts.,Verify that vector/document retrieval functions require and filter by caller authorization scope or tenant ID.,multi-tenant or access-controlled RAG systems,"EU AI Act Art 10 | NIST GOVERN-6, MANAGE-3 | OWASP LLM02, LLM08 | ATLAS AML.M0018 | MIT 1.2. Privacy violations"
11
+ GR-02.5,Data governance,"Implement retention schedules, right-to-erasure purge routines, and versioned snapshots for datasets and embeddings.",Provide deterministic delete-by-source-ID and TTL expiration jobs across document stores and caches.,Verify existence of TTL/retention enforcement or document deletion routines in storage modules.,systems persisting user or enterprise data,"EU AI Act Art 10 | NIST GOVERN-1, MANAGE-4 | OWASP LLM02 | MIT 1.2. Privacy violations"
12
+ GR-03.1,Transparency & disclosure,Disclose clearly to users and downstream consumers when they are interacting with an AI system or reading AI-generated content.,"Attach explicit AI disclosure headers, metadata flags, or UI notices to all generated responses.",Verify response envelopes or UI rendering templates include an explicit AI-generated disclosure indicator.,user-facing AI and content generation systems,"EU AI Act Art 13, 50 | NIST GOVERN-4, MAP-1, MEASURE-2 | OWASP LLM09 | MIT 4.3. Lack of transparency or interpretability; 3.1. Pollution of information ecosystem and loss of shared reality"
13
+ GR-03.2,Transparency & disclosure,Ground generated claims in verifiable source citations and cross-check citations against the underlying corpus or files.,Require structured citations on model outputs and deterministically verify file/document existence and line/span bounds.,Verify post-generation citation validation logic that checks cited sources exist and demotes or flags unverified claims.,"RAG, coding, compliance, and analytical AI systems","EU AI Act Art 13 | NIST MEASURE-2, MANAGE-2 | OWASP LLM09 | MIT 3.2. False or misleading information; 4.3. Lack of transparency or interpretability"
14
+ GR-03.3,Transparency & disclosure,"Provide model/system cards documenting model architecture, version, known limitations, and evaluation benchmarks.",Ship structured system/model metadata alongside code and expose version/capability descriptors via CLI or API.,"Verify presence of structured system metadata, version identifiers, and documented limitations.",any AI system,"EU AI Act Art 11, 13, 53 | NIST GOVERN-1, MAP-1, MEASURE-4 | MIT 4.3. Lack of transparency or interpretability"
15
+ GR-03.4,Transparency & disclosure,"Expose interpretable reasoning traces, confidence indicators, or uncertainty warnings on automated assessments.","Include structured reasoning fields, uncertainty/inconclusive states, and audit trails in assessment outputs.",Verify output schemas require explicit reasoning and support inconclusive/uncertainty states rather than forced binary verdicts.,decision-support and analytical AI systems,"EU AI Act Art 13, 14 | NIST MEASURE-2, MANAGE-1 | OWASP LLM09 | MIT 4.3. Lack of transparency or interpretability; 5.1. Overreliance and unsafe use"
16
+ GR-04.1,Human oversight,"Require human-in-the-loop (HITL) approval before executing high-impact, irreversible, or state-mutating actions.","Gate destructive, financial, or external write tool calls behind an explicit confirmation token or approval queue.",Verify that state-mutating or high-risk actions check for explicit user/operator approval before execution.,agentic and tool-using AI systems,"EU AI Act Art 14 | NIST GOVERN-3, MAP-3, MANAGE-2 | OWASP LLM06 | ATLAS AML.M0018 | MIT 5.2. Loss of human agency and autonomy; 7.2. AI possessing dangerous capabilities"
17
+ GR-04.2,Human oversight,"Provide an immediate stop, cancellation, or kill-switch mechanism to halt autonomous loops and background tasks.","Enforce hard iteration caps, timeouts, and signal/cancellation checks inside agent and audit loops.","Verify bounded loop counters (e.g., max revisions/steps) and clean abort handling in execution drivers.",autonomous and multi-step AI pipelines,"EU AI Act Art 14 | NIST GOVERN-3, MANAGE-3 | OWASP LLM06, LLM10 | MIT 7.4. Lack of capability to control AI; 5.2. Loss of human agency and autonomy"
18
+ GR-04.3,Human oversight,"Support human override, correction, and escalation queues for contested or ambiguous model verdicts.",Surface unresolved disputes as explicit open_questions for human review rather than silently masking disagreements.,Verify that unconverged or contested items are collected into an explicit open_questions or escalation structure.,decision-support and auditing AI systems,"EU AI Act Art 14 | NIST GOVERN-3, MANAGE-1, MANAGE-4 | OWASP LLM09 | MIT 5.1. Overreliance and unsafe use"
19
+ GR-04.4,Human oversight,"Prevent automation bias by presenting counter-evidence, adversarial critiques, and verification status alongside conclusions.",Run an independent adversarial critique pass and highlight demoted or weakly supported claims in reports.,Verify presence of adversarial review/critique stages and explicit demotion markers in user-facing reports.,"analytical, compliance, and recommendation AI systems","EU AI Act Art 14 | NIST MAP-3, MEASURE-2 | OWASP LLM09 | MIT 5.1. Overreliance and unsafe use"
20
+ GR-05.1,Robustness & accuracy,Validate all LLM outputs against strict structural and type schemas before downstream consumption or persistence.,Parse and validate every LLM response against JSON Schema or Pydantic models with extra='forbid'; reject malformed payloads.,Verify strict schema validation on LLM responses and final report artifacts prior to writing or acting on them.,all structured-output LLM and agentic systems,"EU AI Act Art 15 | NIST MEASURE-2, MANAGE-2 | OWASP LLM05, LLM09 | ATLAS AML.M0015 | MIT 7.3. Lack of capability or robustness"
21
+ GR-05.2,Robustness & accuracy,Detect and mitigate hallucinations by cross-verifying model assertions against deterministic ground-truth sources.,"Resolve file paths, line numbers, or database keys cited by the LLM and automatically demote unverified claims.",Verify deterministic cross-verification of model citations/facts and automatic demotion + logging on failure.,all generative AI systems making factual or code assertions,"EU AI Act Art 15 | NIST MEASURE-2, MANAGE-2 | OWASP LLM09 | MIT 3.2. False or misleading information; 7.3. Lack of capability or robustness"
22
+ GR-05.3,Robustness & accuracy,"Handle malformed inputs, transient failures, and partial responses gracefully with deterministic error codes and recovery paths.",Wrap file/JSON parsing and phase transitions in explicit error handlers that emit actionable error messages and non-zero exit codes.,Verify deterministic validation error handling and non-zero exit codes when inputs or responses are missing or invalid.,all AI systems,"EU AI Act Art 15 | NIST MEASURE-2, MANAGE-3 | OWASP LLM05, LLM10 | MIT 7.3. Lack of capability or robustness"
23
+ GR-05.4,Robustness & accuracy,"Maintain an automated regression and unit test suite covering pipeline invariants, edge cases, and failure modes.","Ship automated pytest tests verifying happy paths, invalid inputs, hallucination demotions, and schema drift.","Verify existence of automated tests covering schema validation, citation verification, and pipeline phase transitions.",all AI systems,"EU AI Act Art 15, 43 | NIST MEASURE-1, MEASURE-2, MEASURE-4 | MIT 7.3. Lack of capability or robustness"
24
+ GR-05.5,Robustness & accuracy,Enforce multi-agent or adversarial cross-checking to reduce single-pass model blind spots and drift.,"Pair primary assessment phases with bounded adversarial red-team audit phases that can CONFIRM, REVISE, or INVESTIGATE.",Verify implementation of adversarial audit phases that critique and apply deltas to primary LLM outputs.,high-assurance analytical and agentic systems,"EU AI Act Art 15 | NIST MEASURE-2, MANAGE-2 | OWASP LLM09 | ATLAS AML.M0003 | MIT 7.3. Lack of capability or robustness"
25
+ GR-06.1,Adversarial security,Defend against direct and indirect prompt injection by separating untrusted data from system instructions and constraining model outputs.,"Isolate untrusted repository content into structured data fields, forbid prose execution, and require strict JSON schema adherence.",Verify that untrusted inputs are structured separately from instructions and responses are strictly validated as data.,all LLM and agentic systems processing untrusted inputs,"EU AI Act Art 15 | NIST MAP-3, MEASURE-2, MANAGE-3 | OWASP LLM01, LLM08 | ATLAS AML.M0015, AML.M0019 | MIT 2.2. AI system security vulnerabilities and attacks"
26
+ GR-06.2,Adversarial security,"Sanitize and validate all model-generated strings, paths, and parameters before passing them to OS, filesystem, shell, or database APIs.","Resolve paths against project_root, reject '..' traversal, avoid shell=True, and validate identifiers against strict regexes.",Verify path traversal checks (relative_to project_root) and absence of unsanitized shell interpolation.,"all AI systems interacting with filesystems, shells, or APIs","EU AI Act Art 15 | NIST MEASURE-2, MANAGE-3 | OWASP LLM05, LLM06 | ATLAS AML.M0015, AML.M0018 | MIT 2.2. AI system security vulnerabilities and attacks"
27
+ GR-06.3,Adversarial security,"Enforce resource bounds, input size limits, and loop caps to prevent denial-of-service or unbounded token consumption.","Truncate ingested manifest/doc previews (e.g., 6000 chars), cap file walks (e.g., 200 files), and bound audit revisions.",Verify explicit character/file limits on context gathering and hard upper bounds on retry/revision loops.,all LLM and agentic systems,EU AI Act Art 15 | NIST MANAGE-3 | OWASP LLM10 | ATLAS AML.M0004 | MIT 2.2. AI system security vulnerabilities and attacks
28
+ GR-06.4,Adversarial security,Prevent system prompt leakage and sensitive internal context exposure in user-facing or downstream outputs.,Emit only validated contract fields in final reports and agent wrappers; strip internal prompts and raw source excerpts.,"Verify that external output contracts exclude raw system prompts, credentials, and unrestricted source dumps.",all LLM systems,"EU AI Act Art 15 | NIST MEASURE-2, MANAGE-3 | OWASP LLM02, LLM07 | ATLAS AML.M0002, AML.M0015 | MIT 2.2. AI system security vulnerabilities and attacks; 1.2. Privacy violations"
29
+ GR-06.5,Adversarial security,Conduct regular adversarial red-teaming against model verdicts and evidence claims under strict read-scope constraints.,Run dedicated adversary phases constrained to auditing existing citations without inventing new scope.,Verify adversary instructions and delta application forbid inventing new IDs or expanding scope.,all high-assurance AI systems,"EU AI Act Art 15, 55 | NIST MEASURE-2, MANAGE-2 | OWASP LLM01, LLM09 | ATLAS AML.M0003 | MIT 2.2. AI system security vulnerabilities and attacks"
30
+ GR-07.1,Privacy & data protection,Minimize sensitive and personal data exposure by restricting context collection to necessary manifests and bounded file excerpts.,Read only curated manifest filenames with character caps and skip hidden/credential directories during project walks.,Verify that context collection skips dotfiles/environments and enforces strict character and file count limits.,all AI systems accessing user or repository data,"EU AI Act Art 10, 15 | NIST GOVERN-1, MAP-2, MANAGE-3 | OWASP LLM02 | ATLAS AML.M0012 | MIT 1.2. Privacy violations"
31
+ GR-07.2,Privacy & data protection,Prevent unauthorized outbound network exfiltration of codebase or user data during inference and pipeline execution.,Execute assessment pipelines purely locally via disk-queue exchanges without opening outbound network sockets.,Verify that runtime pipeline modules make zero network requests and never import or invoke external HTTP/LLM clients.,all enterprise and local-first AI tools,"EU AI Act Art 15 | NIST GOVERN-6, MANAGE-3 | OWASP LLM02, LLM06 | ATLAS AML.M0018 | MIT 1.2. Privacy violations; 2.2. AI system security vulnerabilities and attacks"
32
+ GR-07.3,Privacy & data protection,"Scrub or exclude secrets, API keys, and credentials from logs, artifacts, and model context windows.",Never read .env or credential files into prompts and avoid hardcoding or logging tokens in pipeline state.,Verify that .env and secret files are excluded from manifest candidate lists and no credentials appear in artifacts.,all AI systems,"EU AI Act Art 10, 15 | NIST MANAGE-3 | OWASP LLM02 | MIT 1.2. Privacy violations"
33
+ GR-07.4,Privacy & data protection,"Confine all runtime state, temporary files, and reports to a designated, git-ignorable workspace directory.","Write all run artifacts, queues, logs, and reports exclusively beneath <project_root>/.armor/.",Verify that all runtime file writes target <project_root>/.armor/ (or atomic temp files in the same directory).,all local and repository-auditing AI tools,"EU AI Act Art 10, 12 | NIST GOVERN-1, MANAGE-3 | OWASP LLM02 | MIT 1.2. Privacy violations"
34
+ GR-08.1,Fairness & non-discrimination,Evaluate and document affected user populations and potential disparate impacts during system profiling.,Require user_population and data_sensitivity fields in the Phase 2 system_profile schema.,Verify that the triage system_profile schema requires explicit user_population profiling.,all AI systems,"EU AI Act Art 10, 13, 27 | NIST MAP-1, MAP-2, MEASURE-2 | MIT 1.1. Unfair discrimination and misrepresentation"
35
+ GR-08.2,Fairness & non-discrimination,"Apply consistent, deterministic scoring weights and evaluation criteria across all assessed components or subjects.","Compute posture scores using a fixed, transparent weight table (VERDICT_WEIGHT) and deterministic sorting.",Verify that coverage and suggestion rankings are computed deterministically from a fixed weight map.,"scoring, ranking, and auditing AI systems","EU AI Act Art 10, 15 | NIST MEASURE-1, MEASURE-2 | MIT 1.1. Unfair discrimination and misrepresentation"
36
+ GR-08.3,Fairness & non-discrimination,Exclude inapplicable criteria from scoring denominators so specialized systems are not unfairly penalized.,Exclude rows with verdict == 'n/a' from in_scope_count and max_possible_points when computing coverage percentages.,Verify that 'n/a' verdicts are excluded from the coverage denominator in report rollup calculations.,evaluation and compliance scoring systems,"EU AI Act Art 9, 15 | NIST MEASURE-1, MEASURE-2 | MIT 1.1. Unfair discrimination and misrepresentation"
37
+ GR-08.4,Fairness & non-discrimination,Audit subjective classifications through an independent review pass to detect biased or unsubstantiated scoping decisions.,Subject all Phase 2 applicability tags (applies/partial_scope/n_a) to the Phase 3 triage-audit red-team loop.,Verify that triage applicability decisions are audited by triage-audit before conformance scoring proceeds.,classification and triage AI systems,"EU AI Act Art 10, 14 | NIST MEASURE-2, MANAGE-2 | MIT 1.1. Unfair discrimination and misrepresentation"
38
+ GR-09.1,Content safety,"Enforce output content quality and forbid vague, boilerplate, or non-actionable safety/remediation advice.",Instruct the model with explicit remediation quality rules and a list of banned boilerplate phrases.,"Verify that conformance instructions mandate codebase-specific, action-leading remediations and list banned phrases.","advisory, code-review, and remediation-generating AI systems","EU AI Act Art 13, 15 | NIST MEASURE-2, MANAGE-2 | OWASP LLM05, LLM09 | MIT 5.1. Overreliance and unsafe use"
39
+ GR-09.2,Content safety,Escape and sanitize untrusted or model-generated strings before rendering them into Markdown tables or reports.,Escape pipe characters ('|') and normalize newlines when rendering model reasoning and actions into Markdown tables.,Verify Markdown rendering functions escape '|' and newlines in table cells.,systems generating Markdown or HTML reports from LLM outputs,EU AI Act Art 15 | NIST MANAGE-3 | OWASP LLM05 | MIT 2.2. AI system security vulnerabilities and attacks
40
+ GR-09.3,Content safety,"Provide safe, deterministic fallback guidance whenever model-generated remediation text is missing or inconclusive.",Supply deterministic fallback action strings for inconclusive or empty-remediation rows during report generation.,Verify report suggestion builder provides explicit fallback action strings when remediation is empty.,advisory and reporting AI systems,"EU AI Act Art 13, 15 | NIST MANAGE-2, MANAGE-3 | OWASP LLM09 | MIT 5.1. Overreliance and unsafe use"
41
+ GR-09.4,Content safety,"Constrain agent wrapper summaries to concise, bounded notes free of raw source code, citations, or prompt instructions.",Enforce a strict <200 character single-line summary note in the agent YAML verdict contract.,Verify agent wrapper specification enforces bounded notes (<200 chars) with no source excerpts or prompts.,agent wrappers and automated CI/CD gates,"EU AI Act Art 13, 15 | NIST GOVERN-4, MANAGE-3 | OWASP LLM02, LLM05 | MIT 1.2. Privacy violations"
42
+ GR-10.1,"Logging, monitoring & incident response","Maintain an immutable, timestamped run directory recording the input/output artifact of every pipeline phase.",Create UTC-timestamped directories under .armor/runs/<iso>/ and persist numbered JSON/MD artifacts for Phases 1–6.,Verify that new_run creates UTC-timestamped run directories and each phase persists its numbered artifact.,all production and auditing AI pipelines,"EU AI Act Art 12, 19 | NIST GOVERN-1, MEASURE-2, MANAGE-4 | ATLAS AML.M0016 | MIT 4.3. Lack of transparency or interpretability"
43
+ GR-10.2,"Logging, monitoring & incident response",Log every detected model hallucination or rejected citation to an append-only structured incident log.,"Append JSONL records with UTC timestamp, phase, sid, citation, and rejection reason to .armor/hallucination_log.jsonl.",Verify that rejected citations in conform and conform-audit are appended to hallucination_log.jsonl.,all citation-backed generative AI systems,"EU AI Act Art 12, 73 | NIST MEASURE-2, MANAGE-4 | OWASP LLM09 | ATLAS AML.M0016 | MIT 3.2. False or misleading information; 4.3. Lack of transparency or interpretability"
44
+ GR-10.3,"Logging, monitoring & incident response","Record full multi-round adversarial audit trails capturing per-round verdict counts, critiques, and before/after row diffs.",Persist audit_trail arrays in 03_triage_audit.json and 05_conformance_audit.json with before/after revision records.,Verify that audit phases record per-round verdict_counts and before/after diffs in audit_trail.,systems with multi-step or adversarial review loops,"EU AI Act Art 12 | NIST MEASURE-2, MANAGE-4 | MIT 4.3. Lack of transparency or interpretability"
45
+ GR-10.4,"Logging, monitoring & incident response",Enforce mandatory skill-usage telemetry tracking across parent skills and delegated sub-agents.,Gate skill execution on emitting start and end telemetry events with propagated PARENT_TRACKERID.,"Verify that SKILL.md defines a mandatory telemetry gate before tool calls, sub-agents, and final responses.",marketplace skills and multi-agent plugins,"EU AI Act Art 12 | NIST GOVERN-1, MANAGE-4 | ATLAS AML.M0016 | MIT 4.3. Lack of transparency or interpretability"
46
+ GR-10.5,"Logging, monitoring & incident response",Expose a deterministic CLI status inspection command reporting live progress across all pipeline phases.,"Support `--status` on the pipeline CLI to inspect whether each phase is pending, awaiting-llm, response-ready, or done.",Verify that `pipeline <project> --status` reports the current run root and the state of every phase.,multi-phase AI pipelines,"EU AI Act Art 12, 14 | NIST GOVERN-3, MANAGE-4 | MIT 4.3. Lack of transparency or interpretability"
47
+ GR-11.1,Supply-chain & component integrity,"Version and content-address all external standards, catalogs, and reference artifacts using cryptographic SHA256 hashes.","Commit catalog and standard updates via content-addressed `<date>__<sha8>` directories, `latest.json`, and `history.jsonl`.",Verify content-addressed SHA256 hashing and manifest/history tracking in ai_standards_versioning.py.,"systems depending on external catalogs, standards, or model artifacts","EU AI Act Art 15, 53 | NIST GOVERN-6, MAP-2, MANAGE-3 | OWASP LLM03 | ATLAS AML.M0007, AML.M0014 | MIT 2.2. AI system security vulnerabilities and attacks"
48
+ GR-11.2,Supply-chain & component integrity,Compute deterministic source fingerprints of the target codebase to detect concurrent edits or environment drift.,Hash sorted relative paths and modification timestamps across tracked source extensions via project_fingerprint().,Verify that project_fingerprint() produces a deterministic SHA256 prefix that changes whenever a source file is edited.,codebase analysis and auditing tools,"EU AI Act Art 15 | NIST MEASURE-2, MANAGE-3 | OWASP LLM03 | ATLAS AML.M0014 | MIT 7.3. Lack of capability or robustness"
49
+ GR-11.3,Supply-chain & component integrity,Minimize third-party runtime dependencies and isolate network-fetching utilities from the core execution runtime.,Require only pydantic at runtime (stdlib for everything else); isolate requests/pyyaml to offline catalog refresh scripts.,Verify that core `armor/` runtime modules import only stdlib and pydantic.,all software and AI plugins,"EU AI Act Art 15 | NIST GOVERN-6, MANAGE-3 | OWASP LLM03 | ATLAS AML.M0014 | MIT 2.2. AI system security vulnerabilities and attacks"
50
+ GR-11.4,Supply-chain & component integrity,Keep packaged JSON Schemas byte-synchronized with runtime data models and provide automated drift verification.,Ship `posture_report.schema.json` generated from `PostureReport.model_json_schema()` and enforce byte equality in tests.,Verify `--write-schema` and test assertions that `posture_report.schema.json` byte-equals `PostureReport.model_json_schema()`.,all schema-driven AI pipelines,"EU AI Act Art 15 | NIST MEASURE-2, MANAGE-3 | OWASP LLM03, LLM05 | MIT 7.3. Lack of capability or robustness"
51
+ GR-12.1,Agentic autonomy & action safety,Enforce strict phase ordering so downstream agent actions cannot execute before prerequisite artifacts are verified.,"Guard each phase entry point with `require_phase_done(run, prerequisite_phase)` and abort with exact unblock instructions.",Verify that calling later phases before earlier phases raises SystemExit via `require_phase_done`.,multi-step agentic workflows,"EU AI Act Art 14, 15 | NIST GOVERN-3, MANAGE-3 | OWASP LLM06 | ATLAS AML.M0018 | MIT 7.4. Lack of capability to control AI; 5.2. Loss of human agency and autonomy"
52
+ GR-12.2,Agentic autonomy & action safety,Prevent recursive self-auditing or runaway self-targeting in autonomous agent wrappers.,Enforce a mandatory self-targeting guard in `armor-agent` that refuses paths resolving to `armor-agent-plugin`.,Verify that the agent wrapper specification mandates a self-targeting refusal returning `verdict: fail` and zeroed counts.,autonomous auditing and delegation agents,"EU AI Act Art 14, 15 | NIST GOVERN-3, MANAGE-3 | OWASP LLM06, LLM10 | ATLAS AML.M0018 | MIT 7.4. Lack of capability to control AI; 7.1. AI pursuing its own goals in conflict with human goals or values"
53
+ GR-12.3,Agentic autonomy & action safety,Protect concurrent file mutations with cross-platform exclusive file locks and atomic write-and-replace primitives.,"Provide `exclusive_lock`, `locked_write` (mkstemp + os.replace), and `locked_append` in `file_lock.py`.",Verify cross-platform locking (`msvcrt` / `fcntl` + `threading.Lock`) and atomic `os.replace` writes in `file_lock.py`.,concurrent or multi-agent local workflows,EU AI Act Art 15 | NIST MANAGE-3 | OWASP LLM06 | MIT 7.3. Lack of capability or robustness
54
+ GR-12.4,Agentic autonomy & action safety,Constrain red-team adversary agents to read-only verification of cited evidence without expanding scope or inventing findings.,Instruct adversary phases not to read files beyond cited `file:line` spans and ignore any unknown SIDs in apply_deltas.,Verify `apply_deltas` ignores unknown SIDs (recording them in `unknown_sids`) and preserves original row ordering.,multi-agent and adversarial AI systems,"EU AI Act Art 14, 15 | NIST GOVERN-3, MEASURE-2, MANAGE-3 | OWASP LLM06 | ATLAS AML.M0018 | MIT 7.4. Lack of capability to control AI"