@blamejs/exceptd-skills 0.13.62 → 0.13.64

package/data/_indexes/token-budget.json

@@ -3,8 +3,8 @@
     "schema_version": "1.0.0",
     "tokenizer_note": "Character-density approximation: 1 token ≈ 4 chars. This is the canonical rule-of-thumb for OpenAI tokenizers on English+technical text. Claude's tokenizer is typically more efficient on prose; treat this as an upper-bound budget for both. Consumers with stricter precision needs should re-tokenize with their own tokenizer.",
     "approx_chars_per_token": 4,
-    "total_chars": 1671867,
-    "total_approx_tokens": 417973,
+    "total_chars": 1672960,
+    "total_approx_tokens": 418246,
     "skill_count": 42
   },
   "skills": {
@@ -1015,10 +1015,10 @@
     },
     "researcher": {
       "path": "skills/researcher/skill.md",
-      "bytes": 32800,
-      "chars": 32626,
-      "lines": 338,
-      "approx_tokens": 8157,
+      "bytes": 33661,
+      "chars": 33479,
+      "lines": 342,
+      "approx_tokens": 8370,
       "approx_chars_per_token": 4,
       "sections": {
         "frontmatter-scope": {
@@ -1047,9 +1047,9 @@
           "approx_tokens": 377
         },
         "analysis-procedure": {
-          "bytes": 13999,
-          "chars": 13871,
-          "approx_tokens": 3468
+          "bytes": 14860,
+          "chars": 14724,
+          "approx_tokens": 3681
         },
         "output-format": {
           "bytes": 3163,
@@ -1270,10 +1270,10 @@
     },
     "defensive-countermeasure-mapping": {
       "path": "skills/defensive-countermeasure-mapping/skill.md",
-      "bytes": 33463,
-      "chars": 33325,
+      "bytes": 33707,
+      "chars": 33565,
       "lines": 304,
-      "approx_tokens": 8331,
+      "approx_tokens": 8391,
       "approx_chars_per_token": 4,
       "sections": {
         "threat-context": {
@@ -1282,9 +1282,9 @@
           "approx_tokens": 882
         },
         "framework-lag-declaration": {
-          "bytes": 4420,
-          "chars": 4404,
-          "approx_tokens": 1101
+          "bytes": 4556,
+          "chars": 4536,
+          "approx_tokens": 1134
         },
         "ttp-mapping": {
           "bytes": 1957,
@@ -1312,9 +1312,9 @@
           "approx_tokens": 694
         },
         "defensive-countermeasure-mapping": {
-          "bytes": 3321,
-          "chars": 3295,
-          "approx_tokens": 824
+          "bytes": 3429,
+          "chars": 3403,
+          "approx_tokens": 851
         }
       }
     },

package/data/atlas-ttps.json

@@ -512,7 +512,7 @@
       "Multiple production AI assistant prompt injection incidents 2025-2026"
     ],
     "framework_gap": true,
-    "framework_gap_detail": "No framework has a control for prompt injection as an access control failure vector. The attack uses the AI service account's authorized permissions — from AC-2's perspective, the access is authorized. MITRE ATLAS v5.1.0 documents the technique; no framework has implemented controls. OWASP LLM Top 10 documents the class; it is not incorporated in any compliance framework.",
+    "framework_gap_detail": "No framework has a control for prompt injection as an access control failure vector. The attack uses the AI service account's authorized permissions — from AC-2's perspective, the access is authorized. MITRE ATLAS v5.6.0 documents the technique; no framework has implemented controls. OWASP LLM Top 10 documents the class; it is not incorporated in any compliance framework.",
     "controls_that_partially_help": [
       "NIST-800-53-AC-2",
       "ISO-27001-2022-A.8.28"

package/data/framework-control-gaps.json

@@ -365,7 +365,7 @@
       "Treating 'Top 25 addressed' as a compliance signal creates a compliance-theatre risk for organisations with significant AI surface",
       "No cross-walk requirement to ATLAS TTPs — CWE addresses weaknesses; ATLAS addresses adversary techniques. Both are needed for AI coverage"
     ],
-    "real_requirement": "Programmes that claim 'Top 25 addressed' as compliance evidence must additionally: (1) enumerate AI-relevant CWEs outside the Top 25 (CWE-1426 Improper Output Validation, CWE-1039 Inadequate Detection of Adversarial Input, CWE-1230 Exposure of Sensitive Info Through Metadata) with explicit treatment, (2) cross-walk to ATLAS v5.1.0 TTPs for adversarial coverage, (3) re-baseline against the next-published Top 25 with delta analysis. Aligns with EU CRA Annex I, UK NCSC, AU ISM, ISO 27001 A.8.28.",
+    "real_requirement": "Programmes that claim 'Top 25 addressed' as compliance evidence must additionally: (1) enumerate AI-relevant CWEs outside the Top 25 (CWE-1426 Improper Output Validation, CWE-1039 Inadequate Detection of Adversarial Input, CWE-1230 Exposure of Sensitive Info Through Metadata) with explicit treatment, (2) cross-walk to ATLAS v5.6.0 TTPs for adversarial coverage, (3) re-baseline against the next-published Top 25 with delta analysis. Aligns with EU CRA Annex I, UK NCSC, AU ISM, ISO 27001 A.8.28.",
     "status": "open",
     "opened_date": "2026-05-11",
     "evidence_cves": [],
@@ -1586,7 +1586,7 @@
       "LLM-API-as-C2 (SesameOp pattern, ATLAS AML.T0096) is not in the clause 6.1.2 example threat list — risk register templates omit it",
       "No requirement to link AI risk register entries to specific TTP IDs (ATLAS / ATT&CK) — risks remain framework-internal abstractions"
     ],
-    "real_requirement": "Clause 6.1.2 risk registers must (1) ingest ATLAS v5.1.0 TTPs as enumerated AI-specific threat sources, (2) cross-reference jurisdictional obligations (EU AI Act Annex III, NIS2 Art. 21, DORA Art. 28, UK CAF B4, AU ISM AI annex, ISO 27001:2022 A.5.7), (3) include AI-API-as-C2 and prompt-injection-as-RCE as named scenarios, (4) be re-run on threat-intel triggers, not only on calendar cycles.",
+    "real_requirement": "Clause 6.1.2 risk registers must (1) ingest ATLAS v5.6.0 TTPs as enumerated AI-specific threat sources, (2) cross-reference jurisdictional obligations (EU AI Act Annex III, NIS2 Art. 21, DORA Art. 28, UK CAF B4, AU ISM AI annex, ISO 27001:2022 A.5.7), (3) include AI-API-as-C2 and prompt-injection-as-RCE as named scenarios, (4) be re-run on threat-intel triggers, not only on calendar cycles.",
     "status": "open",
     "opened_date": "2026-05-11",
     "evidence_cves": [],
@@ -6199,7 +6199,7 @@
     }
   },
   "ATLAS-AML.T0048": {
-    "framework": "MITRE ATLAS v5.1.0",
+    "framework": "MITRE ATLAS v5.6.0",
     "control_id": "AML.T0048",
     "control_name": "External Harms — ML Supply Chain Compromise (bundled-codec / inference-server class)",
     "designed_for": "ATLAS AML.T0048 catalogues external harms from ML supply-chain compromise, including malicious model weights, poisoned training data, and compromised ML libraries. The technique-level guidance covers detection and mitigation at the model-artifact and library-consumption layer.",