actionwall 0.1.0 → 0.2.0

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.
@@ -1,4 +1,4 @@
1
- # ActionWall — Architecture
1
+ # ActionWall: Architecture
2
2
 
3
3
  ## 1. Bird's-eye view
4
4
 
@@ -22,20 +22,20 @@ ActionWall is a **library first**. The engine is a pure(ish) function: `(Action,
22
22
 
23
23
  ## 2. Data model (src/core/types.ts)
24
24
 
25
- - **Action** — what the agent wants: `{ id, timestamp, agentId, tool, params, context }`.
25
+ - **Action**: what the agent wants: `{ id, timestamp, agentId, tool, params, context }`.
26
26
  `tool` is a canonical name (`shell`, `file.read`, `http.request`, …). `context.toolOutput` carries untrusted external content for injection scanning.
27
- - **Finding** — one violation: `{ scanner, ruleId, severity, title, reason, policyRef, evidence }`. Evidence is always redacted and truncated.
28
- - **Decision** — `{ verdict, risk, findings, summary, latencyMs, timestamp }`.
29
- - **Verdict mapping** — policy `risk_thresholds`: severities in `block` → BLOCK, in `ask` → ASK_HUMAN, otherwise ALLOW.
27
+ - **Finding**: one violation: `{ scanner, ruleId, severity, title, reason, policyRef, evidence }`. Evidence is always redacted and truncated.
28
+ - **Decision**: `{ verdict, risk, findings, summary, latencyMs, timestamp }`.
29
+ - **Verdict mapping**: policy `risk_thresholds`: severities in `block` → BLOCK, in `ask` → ASK_HUMAN, otherwise ALLOW.
30
30
 
31
31
  ## 3. The pipeline (src/core/engine.ts)
32
32
 
33
- 1. Run every scanner in order: toolPermissions → commandRisk → filePolicy → secrets → urlPolicy → injection. (Permissions first — the cheap coarse gate; content scanners next.)
33
+ 1. Run every scanner in order: toolPermissions → commandRisk → filePolicy → secrets → urlPolicy → injection. (Permissions first, the cheap coarse gate; content scanners next.)
34
34
  2. Aggregate: decision risk = max finding severity (LOW if none).
35
35
  3. Map through `risk_thresholds`.
36
36
  4. Append to the audit log (unless disabled or ALLOW with `log_allowed: false`).
37
37
 
38
- **Failure semantics.** A scanner that throws produces a synthetic finding `SCANNER-FAILURE` — CRITICAL (block) when `fail_closed: true` (default), LOW when fail-open. An audit write failure does not lose the decision; it appends an `AUDIT-WRITE-FAILURE` finding. Rationale: a firewall that crashes open is worse than no firewall, but operators must be able to opt out explicitly for availability-critical paths.
38
+ **Failure semantics.** A scanner that throws produces a synthetic finding `SCANNER-FAILURE`, CRITICAL (block) when `fail_closed: true` (default), LOW when fail-open. An audit write failure does not lose the decision; it appends an `AUDIT-WRITE-FAILURE` finding. Rationale: a firewall that crashes open is worse than no firewall, but operators must be able to opt out explicitly for availability-critical paths.
39
39
 
40
40
  ## 4. Scanner contract
41
41
 
@@ -47,10 +47,10 @@ interface Scanner {
47
47
  ```
48
48
 
49
49
  Rules for implementations:
50
- - **Pure** — no command execution, no network. Read-only filesystem access is allowed where the scanner's purpose demands it (bulk-delete counting) and must be bounded.
50
+ - **Pure**: no command execution, no network. Read-only filesystem access is allowed where the scanner's purpose demands it (bulk-delete counting) and must be bounded.
51
51
  - **Return []** for anything not applicable (e.g. commandRisk on an http.request).
52
- - **Redact evidence** — anything resembling a credential passes through `redactText()`.
53
- - **Cite policy** — findings set `policyRef` so the report can point at the exact knob.
52
+ - **Redact evidence**: anything resembling a credential passes through `redactText()`.
53
+ - **Cite policy**: findings set `policyRef` so the report can point at the exact knob.
54
54
 
55
55
  ### Adding a scanner
56
56
 
@@ -58,32 +58,32 @@ Rules for implementations:
58
58
  2. Add it to `DEFAULT_SCANNERS` in `src/core/engine.ts` (or inject via `new ShieldEngine(policy, { scanners })`).
59
59
  3. Add tests + a row in REQUIREMENTS/README.
60
60
 
61
- This is also the extension seam for ML-based classifiers (e.g. a small model scoring injection likelihood) — they return findings like any rule-based scanner.
61
+ This is also the extension seam for ML-based classifiers (e.g. a small model scoring injection likelihood); they return findings like any rule-based scanner.
62
62
 
63
63
  ## 5. Key design decisions
64
64
 
65
- ### D1 — Rule-based first, ML later
65
+ ### D1: Rule-based first, ML later
66
66
  Rules are deterministic, testable, fast (~1–5 ms), auditable, and explainable. For a *firewall* (a control-plane component), "why was this blocked?" must have an exact answer. ML classifiers are a great **addition** (pluggable scanner) but a poor foundation: non-deterministic, unexplainable, and gameable.
67
67
 
68
- ### D2 — Analyze, never execute
69
- The shell tokenizer (`util/shell.ts`) is quote-aware and handles chains (`&&`, `;`, `||`, `|`) so each segment is inspected; a fork bomb is matched on the full command because it spans operators. We deliberately do not resolve commands "the way bash would" beyond tokenization — deeper emulation adds complexity and new bypasses; the residual ambiguity is handled by conservative rules (e.g. command substitution = MEDIUM, eval = HIGH).
68
+ ### D2: Analyze, never execute
69
+ The shell tokenizer (`util/shell.ts`) is quote-aware and handles chains (`&&`, `;`, `||`, `|`) so each segment is inspected; a fork bomb is matched on the full command because it spans operators. We deliberately do not resolve commands "the way bash would" beyond tokenization: deeper emulation adds complexity and new bypasses; the residual ambiguity is handled by conservative rules (e.g. command substitution = MEDIUM, eval = HIGH).
70
70
 
71
- ### D3 — Count, don't guess, for bulk deletes
71
+ ### D3: Count, don't guess, for bulk deletes
72
72
  The headline policy ("no more than 10 automatic deletions") is enforced by **actually counting** the files a glob would remove via a bounded, read-only walk (`util/paths.ts`: 5000-entry cap, depth 12, glob expansion). Severity escalates to CRITICAL when the count exceeds the limit. This turns an intention-based rule ("destructive") into a measurable one ("12 > 10").
73
73
 
74
- ### D4 — Path normalization against traversal & platform drift
75
- All paths are home-expanded, resolved, and normalized to POSIX form with any Windows drive prefix stripped before matching (`/etc/**` and `~/.ssh/**` work everywhere). Because matching happens on the *resolved* path, `../../` traversal is defeated by construction — the resolved location is what gets policy-checked. Protected-path rules are first-match-wins so ordering is explicit in the policy file.
74
+ ### D4: Path normalization against traversal & platform drift
75
+ All paths are home-expanded, resolved, and normalized to POSIX form with any Windows drive prefix stripped before matching (`/etc/**` and `~/.ssh/**` work everywhere). Because matching happens on the *resolved* path, `../../` traversal is defeated by construction: the resolved location is what gets policy-checked. Protected-path rules are first-match-wins so ordering is explicit in the policy file.
76
76
 
77
- ### D5 — Secrets: signatures + entropy + redaction everywhere
78
- Well-known formats (AWS, GitHub, OpenAI, Stripe, Slack, JWT, PEM blocks) are regex signatures; unknown formats are covered by a Shannon-entropy heuristic (≥ 32 chars, entropy ≥ 4.0 → MEDIUM). Placeholder values (`changeme`, `${DB_PASS}`) are filtered to avoid alert fatigue. Critically, the same pattern set drives **log redaction** — the audit trail can't store what the scanner would have blocked.
77
+ ### D5: Secrets with signatures, entropy and redaction everywhere
78
+ Well-known formats (AWS, GitHub, OpenAI, Stripe, Slack, JWT, PEM blocks) are regex signatures; unknown formats are covered by a Shannon-entropy heuristic (≥ 32 chars, entropy ≥ 4.0 → MEDIUM). Placeholder values (`changeme`, `${DB_PASS}`) are filtered to avoid alert fatigue. Critically, the same pattern set drives **log redaction**: the audit trail can't store what the scanner would have blocked.
79
79
 
80
- ### D6 — SSRF guard with a metadata-endpoint special case
80
+ ### D6: SSRF guard with a metadata-endpoint special case
81
81
  `169.254.169.254` (and GCP/Aliyun equivalents) get CRITICAL because that's the canonical cloud-credential-theft path for compromised agents; the rest of RFC1918/loopback/link-local gets HIGH. Domain allow/deny lists support `*.example.com` wildcards; an empty allowlist means "any public domain".
82
82
 
83
- ### D7 — Tamper-evident audit, not tamper-proof
84
- JSONL records hash-chain to their predecessor (`hash = sha256(prevHash + record)`); `log --verify` re-walks the chain and names the first broken entry. A local attacker with filesystem write access can still rewrite the *whole* file consistently (they'd need to recompute every hash — feasible), so this is **tamper-evident**, not tamper-proof. Full integrity would need an external anchor (append-only remote sink, signing key in an HSM) — see ROADMAP.
83
+ ### D7: Tamper-evident audit, not tamper-proof
84
+ JSONL records hash-chain to their predecessor (`hash = sha256(prevHash + record)`); `log --verify` re-walks the chain and names the first broken entry. A local attacker with filesystem write access can still rewrite the *whole* file consistently (they'd need to recompute every hash, which is feasible), so this is **tamper-evident**, not tamper-proof. Full integrity would need an external anchor (append-only remote sink, signing key in an HSM). See ROADMAP.
85
85
 
86
- ### D8 — Fail-closed defaults, explicit escapes
86
+ ### D8: Fail-closed defaults, explicit escapes
87
87
  Deny wins over allow in tool permissions; deny-by-default is one flag away; scanner crashes block. Everything that opens the firewall (`fail_open`, `allow_outside_workspace`, `log_allowed: false`) is an explicit, greppable policy decision.
88
88
 
89
89
  ## 6. Latency budget
@@ -1,6 +1,6 @@
1
- # ActionWall — Competitive Landscape
1
+ # ActionWall: Competitive Landscape
2
2
 
3
- > Research snapshot: October 2026. Refresh before launch — this space moves fast (see the [GitHub `ai-firewall` topic](https://github.com/topics/ai-firewall) for new entrants). Links go to primary sources where possible.
3
+ > Research snapshot: October 2026. Refresh before launch: this space moves fast (see the [GitHub `ai-firewall` topic](https://github.com/topics/ai-firewall) for new entrants). Links go to primary sources where possible.
4
4
 
5
5
  **Purpose:** establish that ActionWall's niche is deliberate, not accidental. The one-line summary:
6
6
 
@@ -13,7 +13,7 @@
13
13
  │ OS sandbox / container (E2B, Firecracker, Docker) │ hard isolation,
14
14
  │ │ coarse grain
15
15
  ├──────────────────────────────────────────────────────────────┤
16
- │ ★ ACTION POLICY LAYER — ActionWall lives here │ deterministic,
16
+ │ ★ ACTION POLICY LAYER · ActionWall lives here │ deterministic,
17
17
  │ (tool calls, commands, file ops, URLs, secrets) │ per-action,
18
18
  │ │ explainable
19
19
  ├──────────────────────────────────────────────────────────────┤
@@ -24,22 +24,22 @@
24
24
  └──────────────────────────────────────────────────────────────┘
25
25
  ```
26
26
 
27
- Most public effort has gone into the *bottom* layers (scanning text) and the *top* layer (sandboxing). The middle layer — deterministic per-action policy — is where ActionWall plays, and it composes with the others rather than replacing them.
27
+ Most public effort has gone into the *bottom* layers (scanning text) and the *top* layer (sandboxing). The middle layer, deterministic per-action policy, is where ActionWall plays, and it composes with the others rather than replacing them.
28
28
 
29
29
  ## Comparison table
30
30
 
31
31
  | Project | Shape | What it inspects | Strength | Gap ActionWall fills |
32
32
  |---|---|---|---|---|
33
33
  | **Meta Llama Guard / Prompt Guard 2** | model-based classifier | prompts & responses (text) | industry-standard content classification; fast small models | probabilistic, not explainable; benchmarked weak on *indirect* injection ([promptfoo's finding](https://www.promptfoo.dev/lm-security-db/vuln/llm-guardrail-benchmark-lies-10b59e81/)); no tool-action policy, no audit trail, requires model hosting |
34
- | **Rebuff** | detection framework | prompt text (heuristics + LLM + canary tokens + vector DB) | layered injection detection, canary-token exfil proofing | Python, LLM-dependent; text-level only — says nothing about what the agent *does* next |
34
+ | **Rebuff** | detection framework | prompt text (heuristics + LLM + canary tokens + vector DB) | layered injection detection, canary-token exfil proofing | Python, LLM-dependent; text-level only: says nothing about what the agent *does* next |
35
35
  | **NVIDIA NeMo Guardrails** | guardrail orchestration (Colang) | conversation flows; can wrap other detectors | mature, flexible rails, big ecosystem | dialog-centric; Python; policies are conversational, not per-tool-action with policy files + exit codes |
36
36
  | **Meta LlamaFirewall** | agent security framework | multi-step agent workflows, content | major-vendor backing, broad scope | Python-first; heavier footprint; not an embeddable TS library with a CLI/audit-verify story |
37
37
  | **Agent-Wall** (MCP) | MCP client↔server proxy | tool calls passing through MCP | drop-in protection for MCP agents without code changes | deployment = infrastructure proxy, MCP-transport-scoped; ActionWall is an embeddable library + CLI across *all* tool shapes (shell, file.*, http.*) with filesystem-measured rules |
38
- | **@goplus/agentguard** (npm, v1.2.1, active — 27 releases) | TS guard library + Claude Code skills / MCP hooks | dangerous commands, data leaks, secrets; "20 detection rules, runtime action evaluation, trust registry" | closest npm overlap; actively maintained; backed by GoPlus (a web3 security company); MIT | web3/vendor-flavored (ships axios network calls, a trust registry, Claude-Code/MCP orientation); ActionWall's differentiation to defend: offline & dependency-light, framework-agnostic `(Action, Policy) → Decision` API, filesystem-measured bulk-delete, policy-as-code with zod + CI exit codes, hash-chained audit with `--verify` |
39
- | **agentgate** (npm, v0.16.0, ISC) | agent API gateway | reads execute immediately, **writes queue for human approval** | independently arrived at the ASK_HUMAN idea — external validation that human-in-the-loop is the right primitive | gateway-shaped (transport level, its own client model), not an embeddable content-aware policy engine with shell/file/secret scanners |
38
+ | **@goplus/agentguard** (npm, v1.2.1, active, 27 releases) | TS guard library + Claude Code skills / MCP hooks | dangerous commands, data leaks, secrets; "20 detection rules, runtime action evaluation, trust registry" | closest npm overlap; actively maintained; backed by GoPlus (a web3 security company); MIT | web3/vendor-flavored (ships axios network calls, a trust registry, Claude-Code/MCP orientation); ActionWall's differentiation to defend: offline & dependency-light, framework-agnostic `(Action, Policy) → Decision` API, filesystem-measured bulk-delete, policy-as-code with zod + CI exit codes, hash-chained audit with `--verify` |
39
+ | **agentgate** (npm, v0.16.0, ISC) | agent API gateway | reads execute immediately, **writes queue for human approval** | independently arrived at the ASK_HUMAN idea: external validation that human-in-the-loop is the right primitive | gateway-shaped (transport level, its own client model), not an embeddable content-aware policy engine with shell/file/secret scanners |
40
40
  | **agent-firewall** (npm) | agent↔LLM proxy | loop detection, prompt caching, budget caps | operational cost/behavior controls | different problem (cost & runaway loops), not security policy on tool actions |
41
- | **promptfoo** | eval / red-team harness | your entire guardrail stack | perfect complement: adversarial testing of guardrails in CI | not a runtime control — pair it: promptfoo tests ActionWall's policies |
42
- | **E2B / container sandboxes** | execution isolation | everything inside the sandbox | hard guarantees, real isolation | coarse grain — no explainable per-action verdicts, no policy files, heavy ops; ActionWall complements (see THREAT_MODEL deployment stack) |
41
+ | **promptfoo** | eval / red-team harness | your entire guardrail stack | perfect complement: adversarial testing of guardrails in CI | not a runtime control; pair it: promptfoo tests ActionWall's policies |
42
+ | **E2B / container sandboxes** | execution isolation | everything inside the sandbox | hard guarantees, real isolation | coarse grain: no explainable per-action verdicts, no policy files, heavy ops; ActionWall complements (see THREAT_MODEL deployment stack) |
43
43
  | **Framework-native approvals** (Claude Code permission prompts, etc.) | UX-level prompts | each tool call | zero setup, human in the loop | not policy-as-code, no programmatic audit export, per-app behavior; ActionWall makes the same idea deterministic, configurable, and logged |
44
44
 
45
45
  ## Positioning statement (for README/launch)
@@ -48,27 +48,27 @@ ActionWall is **not** another prompt-injection scanner and **not** a sandbox. It
48
48
 
49
49
  ## Key differentiators
50
50
 
51
- 1. **Action-level, deterministic, explainable** — every block cites a rule ID and the exact policy knob (`file_policy.max_bulk_delete`), unlike classifier verdicts.
52
- 2. **Measured blast radius** — `rm -rf <glob>` is evaluated by counting real files (bounded walk), turning "destructive" from an opinion into a number.
53
- 3. **TypeScript-first** — most guardrail frameworks are Python-first; the JS agent ecosystem (LangChain.js, Vercel AI SDK, TS MCP servers) is underserved.
54
- 4. **Policy-as-code** — YAML + zod validation + `actionwall validate` in CI + exit codes built for pipelines.
55
- 5. **Tamper-evident audit** — SHA-256 hash chain with `log --verify`, secrets redacted at write time.
56
- 6. **Honest scope** — documented limitations (THREAT_MODEL §6) instead of "prevents prompt injection" claims.
51
+ 1. **Action-level, deterministic, explainable**: every block cites a rule ID and the exact policy knob (`file_policy.max_bulk_delete`), unlike classifier verdicts.
52
+ 2. **Measured blast radius**: `rm -rf <glob>` is evaluated by counting real files (bounded walk), turning "destructive" from an opinion into a number.
53
+ 3. **TypeScript-first**: most guardrail frameworks are Python-first; the JS agent ecosystem (LangChain.js, Vercel AI SDK, TS MCP servers) is underserved.
54
+ 4. **Policy-as-code**: YAML + zod validation + `actionwall validate` in CI + exit codes built for pipelines.
55
+ 5. **Tamper-evident audit**: SHA-256 hash chain with `log --verify`, secrets redacted at write time.
56
+ 6. **Honest scope**: documented limitations (THREAT_MODEL §6) instead of "prevents prompt injection" claims.
57
57
 
58
58
  ## Key sources
59
59
 
60
- - [GitHub topic: ai-firewall](https://github.com/topics/ai-firewall) — running index of AI security gateways (HTTP + MCP)
61
- - [promptfoo LLM Security DB: guardrail benchmark limitations](https://www.promptfoo.dev/lm-security-db/vuln/llm-guardrail-benchmark-lies-10b59e81/) — indirect-injection failures in Llama Guard 3 / Prompt Guard 2
62
- - [promptfoo: testing guardrails guide](https://www.promptfoo.dev/docs/guides/testing-guardrails/) — methodology for benchmarking guardrails (missed attacks vs false positives)
63
- - [arXiv 2502.15427: benchmarking guardrails against prompt injection](https://arxiv.org/html/2502.15427v1) — detection gains vs false-positive costs for NeMo-style rails
64
- - [Snyk: Nx malicious package weaponizing AI coding agents](https://snyk.io/blog/weaponizing-ai-coding-agents-for-malware-in-the-nx-malicious-package/) — the incident class that motivates action-level controls
65
- - [GMI Cloud: guardrails in production](https://www.gmicloud.ai/en/blog/llm-guardrails-in-production-content-filtering-jailbreak-detection-and-why-guardrails-alone-are-not-enough) — production pattern of stacking a fast classifier under an orchestration layer
60
+ - [GitHub topic: ai-firewall](https://github.com/topics/ai-firewall), a running index of AI security gateways (HTTP + MCP)
61
+ - [promptfoo LLM Security DB: guardrail benchmark limitations](https://www.promptfoo.dev/lm-security-db/vuln/llm-guardrail-benchmark-lies-10b59e81/), the indirect-injection failures in Llama Guard 3 / Prompt Guard 2
62
+ - [promptfoo: testing guardrails guide](https://www.promptfoo.dev/docs/guides/testing-guardrails/), the methodology for benchmarking guardrails (missed attacks vs false positives)
63
+ - [arXiv 2502.15427: benchmarking guardrails against prompt injection](https://arxiv.org/html/2502.15427v1), detection gains vs false-positive costs for NeMo-style rails
64
+ - [Snyk: Nx malicious package weaponizing AI coding agents](https://snyk.io/blog/weaponizing-ai-coding-agents-for-malware-in-the-nx-malicious-package/), the incident class that motivates action-level controls
65
+ - [GMI Cloud: guardrails in production](https://www.gmicloud.ai/en/blog/llm-guardrails-in-production-content-filtering-jailbreak-detection-and-why-guardrails-alone-are-not-enough), the production pattern of stacking a fast classifier under an orchestration layer
66
66
 
67
67
  ## Evaluating the landscape: key questions, answered (verified against the npm registry, October 2026)
68
68
 
69
69
  Answers to the standard competitor-evaluation questions, with the evidence basis noted. Claims about other packages come from their npm metadata and published descriptions, not source audits.
70
70
 
71
- **What do they do?** The space splits into five shapes: *text classifiers* (Llama Guard, Prompt Guard 2 — score the words going into and out of the model), *detection frameworks* (Rebuff — layered injection detection with canary tokens), *conversation frameworks* (NeMo Guardrails — Colang rules around dialog), *proxies/gateways* (agent-firewall for cost/loops, agentgate for write approvals, Agent-Wall for MCP transport), and *sandboxes* (E2B, containers — isolate execution). Adjacent-but-different: `agentshield` on npm is a hardlink *backup/rollback* tool (acts after the damage), and promptfoo is an eval harness, not a control.
71
+ **What do they do?** The space splits into five shapes: *text classifiers* (Llama Guard, Prompt Guard 2: score the words going into and out of the model), *detection frameworks* (Rebuff: layered injection detection with canary tokens), *conversation frameworks* (NeMo Guardrails: Colang rules around dialog), *proxies/gateways* (agent-firewall for cost/loops, agentgate for write approvals, Agent-Wall for MCP transport), and *sandboxes* (E2B, containers: isolate execution). Adjacent-but-different: `agentshield` on npm is a hardlink *backup/rollback* tool (acts after the damage), and promptfoo is an eval harness, not a control.
72
72
 
73
73
  **What don't they do?** No verified package combines: an in-process deterministic verdict on the individual tool action + a filesystem-*measured* blast radius (counting what a delete would remove) + policy-as-code validated at load + a hash-chained, verifiable audit log + a TypeScript-first API. Each tool covers one layer; the action-policy layer with measurement is the open niche.
74
74
 
@@ -82,18 +82,18 @@ Answers to the standard competitor-evaluation questions, with the evidence basis
82
82
 
83
83
  **Do they protect shell commands?** @goplus/agentguard blocks "dangerous commands" (20 rules, per its description). ActionWall's depth: quote-aware tokenization, chain/pipeline-aware analysis, ~15 rule families (reverse shells, fork bombs, obfuscation, raw device writes…), each finding citing a rule ID and the exact policy knob.
84
84
 
85
- **Do they address prompt injection?** The classifier tier is crowded (Llama Guard, Prompt Guard, Rebuff, Lakera…) and benchmarked weak on *indirect* injection. ActionWall doesn't compete there: it detects common textual patterns and — the real defense — constrains what an injected agent can *do*. Even a perfect classifier can't stop a hallucinated `rm -rf`; someone must judge the action.
85
+ **Do they address prompt injection?** The classifier tier is crowded (Llama Guard, Prompt Guard, Rebuff, Lakera…) and benchmarked weak on *indirect* injection. ActionWall doesn't compete there: it detects common textual patterns and, the real defense, constrains what an injected agent can *do*. Even a perfect classifier can't stop a hallucinated `rm -rf`; someone must judge the action.
86
86
 
87
87
  **Do they have good TypeScript support?** The systemic gap: the major frameworks are Python-first. The JS/TS agent ecosystem (LangChain.js, Vercel AI SDK, TS MCP servers) has essentially one TS-native option (@goplus/agentguard, web3-oriented). This is the clearest market gap ActionWall fills.
88
88
 
89
- **Are their APIs cumbersome?** Classifier APIs are async, remote, and metered per call. ActionWall: `engine.check(action)` — synchronous, local, ~1–5 ms, zero network.
89
+ **Are their APIs cumbersome?** Classifier APIs are async, remote, and metered per call. ActionWall: `engine.check(action)`: synchronous, local, ~1–5 ms, zero network.
90
90
 
91
- **Are they actively maintained?** @goplus/agentguard: yes (27 releases). agent-firewall: rapid successive releases (churn risk). The big frameworks: maintained by Meta/NVIDIA. Honest caveat that applies to ActionWall too: single-maintainer projects carry bus-factor risk — mitigated here by a tiny dependency surface and deep docs.
91
+ **Are they actively maintained?** @goplus/agentguard: yes (27 releases). agent-firewall: rapid successive releases (churn risk). The big frameworks: maintained by Meta/NVIDIA. Honest caveat that applies to ActionWall too: single-maintainer projects carry bus-factor risk, mitigated here by a tiny dependency surface and deep docs.
92
92
 
93
- **Do they have good tests?** Directly verifiable only for ActionWall (76 tests covering every rule, tamper detection, CLI exit codes). For competitors, the check is CI badges and repo inspection — a question to ask *them*.
93
+ **Do they have good tests?** Directly verifiable only for ActionWall (76 tests covering every rule, tamper detection, CLI exit codes). For competitors, the check is CI badges and repo inspection, a question to ask *them*.
94
94
 
95
- **Can ActionWall provide a simpler developer experience?** Yes — that is the pitch: deterministic defaults that work with zero config, a sync API, a 30-second `npx actionwall demo`, human-readable verdicts, and a verifiable audit trail. Simple to adopt, honest about limits, measurable when it blocks.
95
+ **Can ActionWall provide a simpler developer experience?** Yes, that is the pitch: deterministic defaults that work with zero config, a sync API, a 30-second `npx actionwall demo`, human-readable verdicts, and a verifiable audit trail. Simple to adopt, honest about limits, measurable when it blocks.
96
96
 
97
97
  ### Naming provenance (npm registry, verified 2026-10-01)
98
98
 
99
- `agentshield` — taken (backup tool, tomsun28) · `agent-firewall` — taken (LLM proxy) · `agentward` — squatted by a placeholder package · `agentgate` — taken (write-approval gateway) · **`actionwall` — free, adopted** · free at time of check: `toolward`, `agentmoat`, `agentcage`.
99
+ `agentshield`: taken (backup tool, tomsun28) · `agent-firewall`: taken (LLM proxy) · `agentward`: squatted by a placeholder package · `agentgate`: taken (write-approval gateway) · **`actionwall`: free, adopted** · free at time of check: `toolward`, `agentmoat`, `agentcage`.
@@ -1,6 +1,6 @@
1
- # ActionWall — Policy Reference
1
+ # ActionWall: Policy Reference
2
2
 
3
- Policies are YAML (or JSON) documents validated with zod at load time. **Every field is optional** — omitted sections fall back to the built-in defaults shown in `policies/default.yaml`. Invalid fields fail loudly (`actionwall validate`).
3
+ Policies are YAML (or JSON) documents validated with zod at load time. **Every field is optional**: omitted sections fall back to the built-in defaults shown in `policies/default.yaml`. Invalid fields fail loudly (`actionwall validate`).
4
4
 
5
5
  ```bash
6
6
  actionwall validate --policy policies/strict.yaml
@@ -12,7 +12,7 @@ actionwall validate --policy policies/strict.yaml
12
12
  |---|---|---|---|
13
13
  | `version` | `1` | `1` | Schema version. |
14
14
  | `name` | string | `"default"` | Policy name, shown by `validate`. |
15
- | `description` | string | — | Free text. |
15
+ | `description` | string | - | Free text. |
16
16
  | `fail_closed` | bool | `true` | Scanner crash ⇒ CRITICAL finding (block). Set `false` for fail-open. |
17
17
  | `risk_thresholds.block` | RiskLevel[] | `[HIGH, CRITICAL]` | Severities that BLOCK. |
18
18
  | `risk_thresholds.ask` | RiskLevel[] | `[MEDIUM]` | Severities that escalate to ASK_HUMAN. Anything else is ALLOW. |
@@ -26,7 +26,7 @@ actionwall validate --policy policies/strict.yaml
26
26
  | `deny` | string[] | `[]` | Denied tools. **Deny always wins over allow.** |
27
27
  | `agents` | map | `{}` | Per-agent override: `agents.<agentId>` may set `default`, `allow`, `deny`. When present, an agent's own lists replace the global ones. |
28
28
 
29
- Example — a junior agent that may only read and search:
29
+ Example: a junior agent that may only read and search:
30
30
 
31
31
  ```yaml
32
32
  tools:
@@ -50,7 +50,7 @@ command_risk:
50
50
  - "kubectl\\s+delete\\s+(?!pod)" # negative lookahead works too
51
51
  ```
52
52
 
53
- Built-in command rules (not configurable individually — remove the scanner if you need to): rootish `rm` targets, bulk deletes, recursive+force deletion, pipe-to-shell, process-substitution exec, privilege escalation (`sudo`/`su`/`doas`), fork bombs, `dd of=/dev/*`, `mkfs*`, power commands, reverse shells (`nc -e`, `/dev/tcp/`, `socat EXEC:`), env dumps, eval/base64-decode obfuscation, command substitution, `chmod 777`, `git push --force`, remote copies, secret-file reads, custom patterns, overlong commands.
53
+ Built-in command rules (not configurable individually; remove the scanner if you need to): rootish `rm` targets, bulk deletes, recursive+force deletion, pipe-to-shell, process-substitution exec, privilege escalation (`sudo`/`su`/`doas`), fork bombs, `dd of=/dev/*`, `mkfs*`, power commands, reverse shells (`nc -e`, `/dev/tcp/`, `socat EXEC:`), env dumps, eval/base64-decode obfuscation, command substitution, `chmod 777`, `git push --force`, remote copies, secret-file reads, custom patterns, overlong commands.
54
54
 
55
55
  ## `file_policy`
56
56
 
@@ -62,11 +62,11 @@ Built-in command rules (not configurable individually — remove the scanner if
62
62
  | `allow_outside_workspace` | bool | `false` | If false: writes/deletes outside the workspace ⇒ HIGH, reads ⇒ MEDIUM. |
63
63
 
64
64
  **Path rule** = `{ path, permissions, description? }` where `permissions` is one of:
65
- - `[deny]` — no access at all
66
- - `[readonly]` — reads allowed, writes/deletes blocked
67
- - `[allow]` — everything allowed (use to punch holes in earlier rules)
65
+ - `[deny]`: no access at all
66
+ - `[readonly]`: reads allowed, writes/deletes blocked
67
+ - `[allow]`: everything allowed (use to punch holes in earlier rules)
68
68
 
69
- **Ordering matters** — first match wins:
69
+ **Ordering matters**: first match wins:
70
70
 
71
71
  ```yaml
72
72
  protected_paths:
@@ -123,4 +123,4 @@ Scans `context.toolOutput` and `context.userPrompt`. Pattern families: instructi
123
123
 
124
124
  ## Full example (strict production policy)
125
125
 
126
- See [`policies/strict.yaml`](../policies/strict.yaml) — deny-by-default tools, egress allowlist, `hits_for_high: 1`, and per-agent restrictions. [`policies/readonly.yaml`](../policies/readonly.yaml) shows the whole-filesystem `**` readonly pattern with deny rules ordered before it.
126
+ See [`policies/strict.yaml`](../policies/strict.yaml): deny-by-default tools, egress allowlist, `hits_for_high: 1`, and per-agent restrictions. [`policies/readonly.yaml`](../policies/readonly.yaml) shows the whole-filesystem `**` readonly pattern with deny rules ordered before it.
@@ -1,10 +1,10 @@
1
- # ActionWall — Requirements Specification
1
+ # ActionWall: Requirements Specification
2
2
 
3
- Version 1.0 · Status: implemented (v0.1.0) · Identified functional and non-functional requirements, each with acceptance criteria mapped to the test suite.
3
+ Version 1.0 · Status: implemented (v0.1.1) · Identified functional and non-functional requirements, each with acceptance criteria mapped to the test suite.
4
4
 
5
5
  ## 1. Problem statement
6
6
 
7
- AI agents execute tools autonomously: shell commands, file writes, HTTP requests, third-party APIs. A compromised prompt (prompt injection), a hallucinated action, or a bug can cause irreversible damage — deleted files, leaked credentials, SSRF into cloud infrastructure. Existing safeguards (model guardrails, provider "safety" layers) are probabilistic and sit *inside* the model. There is no deterministic, auditable, policy-driven control point between the agent and its tools.
7
+ AI agents execute tools autonomously: shell commands, file writes, HTTP requests, third-party APIs. A compromised prompt (prompt injection), a hallucinated action, or a bug can cause irreversible damage: deleted files, leaked credentials, SSRF into cloud infrastructure. Existing safeguards (model guardrails, provider "safety" layers) are probabilistic and sit *inside* the model. There is no deterministic, auditable, policy-driven control point between the agent and its tools.
8
8
 
9
9
  ## 2. Goals & non-goals
10
10
 
@@ -16,7 +16,7 @@ AI agents execute tools autonomously: shell commands, file writes, HTTP requests
16
16
  - G5: Zero network calls and zero command execution inside the firewall itself.
17
17
 
18
18
  **Non-goals (v1)**
19
- - Running as a network proxy (planned — see ROADMAP).
19
+ - Running as a network proxy (planned, see ROADMAP).
20
20
  - Stopping a *determined* adversary with arbitrary obfuscated payloads; ActionWall raises the floor and adds auditability, it is not a sandbox (see THREAT_MODEL).
21
21
  - Replacing OS-level sandboxing (containers, users, seccomp). It complements them.
22
22
 
@@ -26,7 +26,7 @@ AI agents execute tools autonomously: shell commands, file writes, HTTP requests
26
26
 
27
27
  | ID | Requirement | Acceptance criteria |
28
28
  |---|---|---|
29
- | FR-1 | Engine accepts an Action `{tool, params, agentId, context}` and returns a Decision `{verdict, risk, findings[], summary, latencyMs}`. | `engine.test.ts` — clean action yields `ALLOW`, risk `LOW`, latency ≥ 0. |
29
+ | FR-1 | Engine accepts an Action `{tool, params, agentId, context}` and returns a Decision `{verdict, risk, findings[], summary, latencyMs}`. | `engine.test.ts`: clean action yields `ALLOW`, risk `LOW`, latency ≥ 0. |
30
30
  | FR-2 | Verdicts are exactly `ALLOW`, `BLOCK`, `ASK_HUMAN`. | All tests assert one of three values. |
31
31
  | FR-3 | Severity levels are `LOW < MEDIUM < HIGH < CRITICAL`; the decision's risk is the worst finding. | FR-1 + scanner tests. |
32
32
  | FR-4 | Policy maps severities → verdicts (`risk_thresholds.block` / `.ask`). | `engine.test.ts` custom thresholds: HIGH→ASK when configured. |
@@ -39,9 +39,9 @@ AI agents execute tools autonomously: shell commands, file writes, HTTP requests
39
39
  |---|---|---|---|
40
40
  | FR-7 | toolPermissions | Allow/deny lists with `*` wildcards, global or per-agent; deny wins over allow; deny-by-default mode. | `toolPermissions.test.ts`. |
41
41
  | FR-8 | commandRisk | Tokenize shell commands (quote-aware, chain/pipeline aware) without executing; detect destructive FS ops, privilege escalation, pipe-to-shell, reverse shells, fork bombs, raw device writes, env dumps, obfuscation, custom regex patterns. | `commandRisk.test.ts` (13 cases). |
42
- | FR-9 | commandRisk | **Bulk-delete rule:** expand `rm` targets against the real filesystem (read-only) and block when the count exceeds `file_policy.max_bulk_delete`. | `commandRisk.test.ts` — 12 files with limit 10 → CRITICAL, evidence names the count. |
42
+ | FR-9 | commandRisk | **Bulk-delete rule:** expand `rm` targets against the real filesystem (read-only) and block when the count exceeds `file_policy.max_bulk_delete`. | `commandRisk.test.ts`: 12 files with limit 10 → CRITICAL, evidence names the count. |
43
43
  | FR-10 | filePolicy | Protected-path rules (`deny` / `readonly` / `allow`, first match wins) applied to `file.*` tools **and** file-touching shell commands; workspace containment for reads/writes/deletes. | `filePolicy.test.ts`. |
44
- | FR-11 | secrets | Detect known credential formats + credential assignments (with placeholder filtering) + high-entropy tokens in tool args; never emit the raw secret in findings. | `secrets.test.ts` — AWS/GitHub/OpenAI/private-key/JWT cases; redaction assertions. |
44
+ | FR-11 | secrets | Detect known credential formats + credential assignments (with placeholder filtering) + high-entropy tokens in tool args; never emit the raw secret in findings. | `secrets.test.ts`: AWS/GitHub/OpenAI/private-key/JWT cases; redaction assertions. |
45
45
  | FR-12 | urlPolicy | Scheme allowlist, domain allow/deny lists with wildcards, private-network & cloud-metadata SSRF guard; applies to web tools and to URLs inside shell commands. | `urlPolicy.test.ts`. |
46
46
  | FR-13 | injection | Heuristic detection of instruction overrides, fake system markers, prompt-extraction, markdown-image exfiltration, hidden characters, in `toolOutput` and `userPrompt`; N distinct patterns escalate MEDIUM→HIGH. | `injection.test.ts`. |
47
47
 
@@ -49,7 +49,7 @@ AI agents execute tools autonomously: shell commands, file writes, HTTP requests
49
49
 
50
50
  | ID | Requirement | Acceptance criteria |
51
51
  |---|---|---|
52
- | FR-14 | Every decision (including ALLOW, configurable) is appended to daily JSONL files with action params, verdict, risk, findings. | `cli.test.ts` — check writes to `.actionwall/audit/`. |
52
+ | FR-14 | Every decision (including ALLOW, configurable) is appended to daily JSONL files with action params, verdict, risk, findings. | `cli.test.ts`: check writes to `.actionwall/audit/`. |
53
53
  | FR-15 | Records are hash-chained (SHA-256 over prev hash + record); `verify` re-walks the chain and pinpoints the first broken entry. | `audit.test.ts` + `cli.test.ts` tamper case. |
54
54
  | FR-16 | Secrets in logged params are redacted. | `audit.test.ts`. |
55
55
 
@@ -58,7 +58,7 @@ AI agents execute tools autonomously: shell commands, file writes, HTTP requests
58
58
  | ID | Requirement | Acceptance criteria |
59
59
  |---|---|---|
60
60
  | FR-17 | `check <command-or-json>` with `--tool/--params/--context/--agent/--policy/--json/--no-log`. | `cli.test.ts`. |
61
- | FR-18 | Exit codes: 0 allow, 1 block, 3 ask-human, 2 error — CI usable. | `cli.test.ts`. |
61
+ | FR-18 | Exit codes: 0 allow, 1 block, 3 ask-human, 2 error. CI usable. | `cli.test.ts`. |
62
62
  | FR-19 | `validate`, `log [--last N] [--verify]`, `demo`. | `cli.test.ts`. |
63
63
 
64
64
  ## 4. Non-functional requirements
@@ -72,13 +72,13 @@ AI agents execute tools autonomously: shell commands, file writes, HTTP requests
72
72
  | NFR-5 | **Portability:** Linux, macOS, Windows (drive-letter normalization, home expansion). | Tests run on Windows; CI matrix. |
73
73
  | NFR-6 | **Dependency hygiene:** 3 runtime dependencies (commander, yaml, zod), zero for the core engine. | package.json. |
74
74
  | NFR-7 | **Explainability:** every finding carries ruleId, human-readable reason, and the policy reference that triggered it. | All scanner tests. |
75
- | NFR-8 | **Test coverage:** every scanner, the engine, audit, policy loader, and CLI have automated tests (76 as of v0.1.0). | `npm test`. |
75
+ | NFR-8 | **Test coverage:** every scanner, the engine, audit, policy loader, and CLI have automated tests (76 as of v0.1.1). | `npm test`. |
76
76
 
77
77
  ## 5. Personas
78
78
 
79
- - **Agent developer** — integrates ActionWall into a tool loop; wants `guard()` and clear errors.
80
- - **Security/platform engineer** — writes policies and reviews the audit log; wants `validate` and `log --verify`.
81
- - **Operator/agent owner** — receives ASK_HUMAN escalations and approves or denies.
79
+ - **Agent developer**: integrates ActionWall into a tool loop; wants `guard()` and clear errors.
80
+ - **Security/platform engineer**: writes policies and reviews the audit log; wants `validate` and `log --verify`.
81
+ - **Operator/agent owner**: receives ASK_HUMAN escalations and approves or denies.
82
82
 
83
83
  ## 6. Acceptance scenario (product pitch parity)
84
84
 
package/docs/ROADMAP.md CHANGED
@@ -1,4 +1,4 @@
1
- # ActionWall — Roadmap
1
+ # ActionWall: Roadmap
2
2
 
3
3
  ## v0.1 (current)
4
4
 
@@ -7,28 +7,28 @@
7
7
  - Hash-chained JSONL audit log with `--verify`
8
8
  - 76 tests; presets: default / strict / readonly
9
9
 
10
- ## v0.2 — integration depth
10
+ ## v0.2: integration depth
11
11
 
12
- - [ ] **HTTP service mode** (`actionwall serve`): POST /v1/check, per-tenant policies, p99 < 5 ms budget
12
+ - [x] **HTTP service mode** (`actionwall serve`): POST /v1/check with `x-actionwall-verdict` / `x-actionwall-risk` headers, GET /healthz, per-tenant policies via a tenants file (shipped, unreleased)
13
13
  - [ ] **MCP guard**: run ActionWall as an MCP gateway/proxy so Claude Desktop & friends are protected without code changes
14
14
  - [ ] Official middleware packages: OpenAI tool-calling, Anthropic tool-use, LangChain/LangGraph, Vercel AI SDK
15
15
  - [ ] Async audit writer (batched, back-pressured) for high-throughput service mode
16
16
 
17
- ## v0.3 — smarter detection (shadow-first)
17
+ ## v0.3: smarter detection (shadow-first)
18
18
 
19
19
  - [ ] Pluggable ML classifier for prompt injection (returns standard Findings; runs in shadow mode first)
20
20
  - [ ] Per-scanner exceptions/allowlists (`*.pem` allow for a build agent)
21
21
  - [ ] Configurable rule severities (user-tunable rule table in policy)
22
22
  - [ ] Homoglyph/confusable detection with per-locale opt-in
23
23
 
24
- ## v0.4 — operations & trust
24
+ ## v0.4: operations & trust
25
25
 
26
26
  - [ ] Real approvals workflow for ASK_HUMAN (queue, TTL, auto-deny, Slack/webhook)
27
- - [ ] Remote audit sink (S3 Object Lock / syslog) + periodic signed Merkle checkpoints — tamper-*proof* rather than tamper-evident
28
- - [ ] `actionwall replay` — re-run past actions under a new policy ("what would have been blocked?")
27
+ - [ ] Remote audit sink (S3 Object Lock / syslog) + periodic signed Merkle checkpoints, tamper-*proof* rather than tamper-evident
28
+ - [ ] `actionwall replay`: re-run past actions under a new policy ("what would have been blocked?")
29
29
  - [ ] Metrics endpoint (Prometheus): decisions, verdicts, rule hits, latency histograms
30
30
 
31
- ## v1.0 — hardening
31
+ ## v1.0: hardening
32
32
 
33
33
  - [ ] Differential shell parsing (exchange the tokenizer for a battle-tested parser behind the same interface)
34
34
  - [ ] Policy inheritance & org/team composition
@@ -1,4 +1,4 @@
1
- # ActionWall — Threat Model
1
+ # ActionWall: Threat Model
2
2
 
3
3
  Scope: an AI agent (LLM-driven) running on a developer workstation or a server, executing tools autonomously. ActionWall is the enforcement point between the agent and its tools.
4
4
 
@@ -14,10 +14,10 @@ Scope: an AI agent (LLM-driven) running on a developer workstation or a server,
14
14
 
15
15
  ## 2. Adversaries
16
16
 
17
- 1. **Remote attacker via content** — crafts a web page, file, issue, or email whose *content* the agent will read; goal: prompt-inject the agent into harmful actions (OWASP LLM01).
18
- 2. **Compromised dependency/tool** — an MCP server or npm package the agent uses misbehaves or exfiltrates.
19
- 3. **Confused/degraded agent** — not malicious, just wrong: hallucinated paths, over-broad globs, wrong environment (prod vs dev).
20
- 4. **Local attacker with user-level code execution** — can already do most damage directly; ActionWall's audit log still raises forensics cost (but see Limitations L7).
17
+ 1. **Remote attacker via content**: crafts a web page, file, issue, or email whose *content* the agent will read; goal: prompt-inject the agent into harmful actions (OWASP LLM01).
18
+ 2. **Compromised dependency/tool**: an MCP server or npm package the agent uses misbehaves or exfiltrates.
19
+ 3. **Confused/degraded agent**: not malicious, just wrong: hallucinated paths, over-broad globs, wrong environment (prod vs dev).
20
+ 4. **Local attacker with user-level code execution**: can already do most damage directly; ActionWall's audit log still raises forensics cost (but see Limitations L7).
21
21
 
22
22
  ## 3. Attack scenarios → controls
23
23
 
@@ -54,7 +54,7 @@ Scope: an AI agent (LLM-driven) running on a developer workstation or a server,
54
54
 
55
55
  | OWASP item | ActionWall coverage |
56
56
  |---|---|
57
- | LLM01 Prompt Injection | injection scanner (heuristics) + the entire deny/confirm pipeline; **detection ≠ prevention** — see L1 |
57
+ | LLM01 Prompt Injection | injection scanner (heuristics) + the entire deny/confirm pipeline; **detection ≠ prevention**, see L1 |
58
58
  | LLM02 Sensitive Information Disclosure | secrets scanner, protected paths, redacted audit |
59
59
  | LLM03 Supply Chain | tool permission lists; a compromised tool still runs *inside* policy |
60
60
  | LLM04 Data & Model Poisoning | out of scope (training-time) |
@@ -67,14 +67,14 @@ Scope: an AI agent (LLM-driven) running on a developer workstation or a server,
67
67
 
68
68
  ## 6. Honest limitations
69
69
 
70
- - **L1 — Heuristics are bypassable.** Paraphrased, multilingual, or steganographic injections will pass pattern matching. ActionWall's answer is *defense in depth*: even a successful injection still has to get the resulting *action* past the command/file/URL/secret scanners. Planned: pluggable LLM classifier.
71
- - **L2 — Shell emulation is bounded.** The tokenizer doesn't model expansion semantics (`$IFS`, brace expansion, aliases, heredocs). Residual ambiguity is handled conservatively (substitution/eval rules), not perfectly.
72
- - **L3 — Counting is bounded, not exact.** The bulk-delete walk caps at 5000 entries; beyond the cap it's a lower bound (reported as "5000+"), which is fine — the rule only needs "is it over 10".
73
- - **L4 — Race conditions.** A file could appear between counting and deletion. ActionWall is a gate, not a transactional FS layer.
74
- - **L5 — TOCTOU on decisions.** What the agent does *after* an ALLOW (e.g. writes a script, then a later action runs it) is only caught when that later action is checked.
75
- - **L6 — No OS-level guarantees.** This is policy enforcement in process. Pair with containers, users, seccomp for hard isolation.
76
- - **L7 — Audit chain is tamper-evident, not tamper-proof.** A full-file rewrite with recomputed hashes is detectable only with an external anchor (remote sink / signing) — ROADMAP.
77
- - **L8 — Unicode tricks.** Zero-width characters are flagged; homoglyph substitution is not (too noisy for a default rule).
70
+ - **L1: Heuristics are bypassable.** Paraphrased, multilingual, or steganographic injections will pass pattern matching. ActionWall's answer is *defense in depth*: even a successful injection still has to get the resulting *action* past the command/file/URL/secret scanners. Planned: pluggable LLM classifier.
71
+ - **L2: Shell emulation is bounded.** The tokenizer doesn't model expansion semantics (`$IFS`, brace expansion, aliases, heredocs). Residual ambiguity is handled conservatively (substitution/eval rules), not perfectly.
72
+ - **L3: Counting is bounded, not exact.** The bulk-delete walk caps at 5000 entries; beyond the cap it's a lower bound (reported as "5000+"), which is fine, since the rule only needs "is it over 10".
73
+ - **L4: Race conditions.** A file could appear between counting and deletion. ActionWall is a gate, not a transactional FS layer.
74
+ - **L5: TOCTOU on decisions.** What the agent does *after* an ALLOW (e.g. writes a script, then a later action runs it) is only caught when that later action is checked.
75
+ - **L6: No OS-level guarantees.** This is policy enforcement in process. Pair with containers, users, seccomp for hard isolation.
76
+ - **L7: Audit chain is tamper-evident, not tamper-proof.** A full-file rewrite with recomputed hashes is detectable only with an external anchor (remote sink / signing); see ROADMAP.
77
+ - **L8: Unicode tricks.** Zero-width characters are flagged; homoglyph substitution is not (too noisy for a default rule).
78
78
 
79
79
  ## 7. Deployment guidance (defense in depth)
80
80
 
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "actionwall",
3
- "version": "0.1.0",
3
+ "version": "0.2.0",
4
4
  "description": "Security firewall for AI agents — inspect every tool call, command, file access and HTTP request against a policy before it runs.",
5
5
  "type": "module",
6
6
  "license": "MIT",
@@ -25,7 +25,7 @@
25
25
  "bugs": {
26
26
  "url": "https://github.com/uzairr-arif/actionwall/issues"
27
27
  },
28
- "homepage": "https://github.com/uzairr-arif/actionwall#readme",
28
+ "homepage": "https://uzairr-arif.github.io/actionwall/",
29
29
  "files": [
30
30
  "dist",
31
31
  "policies",
@@ -54,14 +54,14 @@
54
54
  "audit-log"
55
55
  ],
56
56
  "dependencies": {
57
- "commander": "^14.0.0",
57
+ "commander": "^15.0.0",
58
58
  "yaml": "^2.6.0",
59
- "zod": "^3.24.1"
59
+ "zod": "^4.6.7"
60
60
  },
61
61
  "devDependencies": {
62
- "@types/node": "^22.10.0",
62
+ "@types/node": "^26.6.5",
63
63
  "tsx": "^4.19.0",
64
- "typescript": "^5.7.0",
65
- "vitest": "^3.0.0"
64
+ "typescript": "^7.0.2",
65
+ "vitest": "^5.0.5"
66
66
  }
67
67
  }