Skip to content

AAK-COMPOSE-002: Two or more skills share an undeclared writable path

Severity
HIGH
Category
Composition (composition)
Scanner
agent_audit_kit/scanners/composition.py
Auto-fixable
No.
SARIF name
CompositionSharedUndeclaredState

What it detects

Two or more skills that load into the same context both reference the same writable filesystem path, and none of them declares a write capability or names that path as an output. An undeclared path both can reach is a channel between them: one writes what the other reads, with no interface either manifest describes and nothing in a per-skill review to notice. This is the ColluSkill shape (arXiv:2608.09732), where the cooperation is what carries the attack rather than any single skill's behaviour. A declared-capability union cannot see this by construction, because the relevant capability is precisely the one nobody declared.

How to fix

Decide whether the shared path is an interface or an accident. If the skills are meant to exchange data, declare it: name the path as an output of the writer and an input of the reader, so a reviewer and this scanner both see the coupling. If they are not, give each skill its own working directory. Do not resolve it by adding a write capability to one manifest and leaving the path shared - that documents the write without documenting the channel.

To confirm the finding is gone, report only this rule:

agent-audit-kit scan . --rules AAK-COMPOSE-002

Limitations

This is a static pass over declared configuration. It cannot resolve tool registration that happens at runtime, so a component added from a registry fetched at start-up is invisible to it, and a chain assembled at runtime from such components will not appear here. Paths are read literally out of skill bodies and adjacent scripts, so a path built at runtime from a variable or an environment lookup is not seen, and two skills sharing a channel under different spellings of the same location will not be matched. It reports co-reference, not a proven write followed by a read.

References

  • OWASP Agentic Top 10: ASI06 (Memory & Context Poisoning)
  • Adversa AI MCP Top 25 (AAK-defined ids): ADV-INJECT-01 (Shell Injection via MCP)
  • Incidents and advisories: ARXIV-2608.09732

Deep dive

This rule is written up in Skill scanning: per-artifact pre-screen and the composition check.


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-COMPOSE-002.md: the page then shows that text, with the registry reference appended under it.