Skip to content

AAK-AGENT-COMPOSE-001: Skill set's composed capability union crosses a risk boundary no single skill requested

Severity
HIGH
Category
Trust Boundaries (trust-boundary)
Scanner
agent_audit_kit/scanners/skill_composition.py
Auto-fixable
No.
SARIF name
SkillSetCapabilityUnionExfil

What it detects

Unlike the AAK-AGENT-TRUST-* / AAK-SKILL-* rules, which inspect one artifact at a time, this rule evaluates the SET of skills that load into one agent context (all SKILL.md under a common container such as .claude/skills/). It computes the union of declared capability across the set (filesystem read/write, network egress by destination, shell execution, credential access, memory write) and flags a union that crosses a configured risk boundary that no single skill in the set requested. The shipped default: a skill that can read the filesystem or credentials, composed with a skill that can egress to a non-allowlisted destination, is an exfiltration path even when every contributing skill is individually clean. That is the cross-skill composition attack ColluSkill (arXiv:2608.09732) measured at a 96.0% average success rate across six per-skill scanners. The finding names which skill contributed which capability and lists every contributor as a related location. The boundary set and egress allowlist are configurable via .aak/composition-boundaries.yaml; the default and its reasoning are in docs/rules/skill-composition.md.

How to fix

Do not load a skill that reads files or credentials into the same context as a skill that can egress to an unaudited destination. Split them into separate agent contexts, pin each egress skill to an allowlisted destination, or remove the network capability from the read-capable set. Review a new skill for the capability it adds to the union, not only in isolation.

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

agent-audit-kit scan . --rules AAK-AGENT-COMPOSE-001

Limitations

Heuristic on DECLARED capability (allowed-tools / capabilities / egress frontmatter). A skill that under-declares its tools, or reaches a capability through an MCP server it does not name, is not measured here; the per-skill scanners and Python taint analysis cover in-body behaviour. It reasons about capability, not data flow, so it flags a possible exfiltration path, not a proven one. A URL that Node, requests and Rust's url crate read differently from curl, httpx and urlsplit, which keep a backslash inside the authority, is judged by the less safe of the two readings.

References

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-AGENT-COMPOSE-001.md: the page then shows that text, with the registry reference appended under it.