knowledge-synthesis
Fail
Audited by Gen Agent Trust Hub on Nov 5, 2026
Risk Level: HIGHPROMPT_INJECTIONPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill includes a 'Codex compatibility note' section containing meta-instructions that attempt to redefine how the agent loads skills and maps tool names. It specifically instructs the agent to ignore standard mode switches and tool requirements.
- [PRIVILEGE_ESCALATION]: The instructions contain a directive stating that skill activation 'authorizes use of the required spawn_agent subagent(s) for that task.' This is a direct attempt to bypass user confirmation or platform-level authorization for spawning new autonomous agents.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from external research files which may contain untrusted content gathered from the web.
- Ingestion points: Reads evidence from
tmp/research/_evidence-{slug}.mdandtmp/research/_sources-{slug}.md. - Boundary markers: The skill lacks explicit delimiters or instructions to treat the content of these files as untrusted data rather than instructions.
- Capability inventory: The agent has the capability to write files to the file system and, per the prompt injection findings, is encouraged to spawn subagents autonomously.
- Sanitization: No sanitization or verification of the content within the research evidence files is performed before it is synthesized into final reports.
Recommendations
- AI detected serious security threats
Audit Metadata