agent-injection-boundary-test
Pass
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: The phrase "ignore previous instructions" appears in
gemini.md. Detailed analysis confirms this is used as an example of an overt attack pattern that testers should identify, not an actual attempt to manipulate the agent's behavior. The operational directives are aligned with the skill's primary purpose of security auditing. - [INDIRECT_PROMPT_INJECTION]: The skill describes the vulnerability surface for indirect prompt injection (RAG documents, tool outputs, etc.) and provides a procedure for evaluation. It advocates for robust defenses such as "capability isolation" and "allow-lists" rather than brittle prompt-based filters.
- [DATA_EXFILTRATION]: The skill contains strict operational rules against using real credentials or exfiltration endpoints. It recommends the use of "synthetic canaries" and offline "replay mode" to prevent any actual data leak during the security testing process.
Audit Metadata