AAK-AST10-001: Skill ships platform manifests whose security metadata disagrees
- Severity
- MEDIUM
- Category
- Agentic Skills (AST10) (
agentic-skill) - Scanner
agent_audit_kit/scanners/agentic_skills.py- Auto-fixable
- No.
- SARIF name
AgenticSkillCrossPlatformMetadataLoss
What it detects
A skill bundle carries more than one platform manifest — SKILL.md frontmatter, skill.json, manifest.json, package.json — and a security-relevant field declared in one is absent from another, or declares something weaker in one than the other: a risk tier that exists on one platform and not the next, a permission list replaced by a blanket true, a signature field that survives in one format only. Whichever platform loads the weaker manifest runs the skill without the control the other declares. This is OWASP AST10 Cross-Platform Reuse, whose first attack scenario is a risk_tier: L3 skill ported to a platform that does not support risk_tier so the warning is silently dropped. Like the composition rules, this is a question no single artifact can answer: each manifest is internally consistent, and the defect exists only in the disagreement between them.
How to fix
Normalise the security fields across every manifest the bundle ships, which is AST10's own mitigation: risk_tier, permissions, egress and signature should say the same thing in each format or the bundle should ship one format. Where a platform has no equivalent for a field, do not drop it silently — record the gap in the bundle so a reviewer sees that the control is unenforced there rather than absent by design. Re-validate the metadata whenever a skill is ported; equivalence between formats is the assumption this risk is named after.
To confirm the finding is gone, report only this rule:
agent-audit-kit scan . --rules AAK-AST10-001
Limitations
Static, and scoped to what a skill bundle declares on disk. It cannot see a resource that is pinned at install time by tooling outside the bundle, and it does not fetch anything, so a pinned reference that now resolves elsewhere looks the same as one that does not. It compares only fields whose loss is unambiguous, and only in the unambiguous direction: an allow-list replaced by a blanket true, or a declaration replaced by nothing. Two differing permission lists are left alone, because deciding which is weaker needs a judgement this rule cannot make without guessing. A bundle shipping one manifest is out of scope by construction, and so is a manifest that does not parse — which matters here, because frontmatter carrying a deserialization tag is rejected by safe parsing and therefore drops out of this comparison entirely. In that case AAK-AST04-001 is the finding that applies.
References
- OWASP Agentic Top 10: ASI04 (Agentic Supply Chain Vulnerabilities)
- OWASP Agentic Skills Top 10: AST10 (Cross-Platform Reuse)
- Adversa AI MCP Top 25 (AAK-defined ids): ADV-SUPPLY-01 (Runtime Package Fetch)
All rules. This page is generated from the rule registry when the documentation is built; to change it, change the rule in agent_audit_kit/rules/builtin.py. To give the rule a hand-written write-up, add docs/rules/AAK-AST10-001.md: the page then shows that text, with the registry reference appended under it.