scaffolding-openai-agents
Warn
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The
calculatefunction tool inSKILL.mduses theeval()function to process mathematical expressions. Because the input to this tool is generated by the AI agent based on potentially untrusted user prompts, it allows for arbitrary Python code execution. Although a code comment suggests usingsafe_evalin production, providing the raweval()implementation remains a significant security risk for developers using the scaffold. - [INDIRECT_PROMPT_INJECTION]: The skill architecture is susceptible to indirect prompt injection as it processes untrusted user data that can influence agent tool-calling and responses.
- Ingestion points: User queries are ingested through the
Runner.runmethod in themain.pyexample and the FastAPI/askendpoint inSKILL.md. - Boundary markers: The provided agent instructions lack explicit delimiters or specific guard instructions to isolate user data from agent logic.
- Capability inventory: Tools defined in
SKILL.mdinclude powerful capabilities such as code execution viaeval()and network operations viahttpx. - Sanitization: The implemented guardrails in
SKILL.mdrely on simple keyword matching (e.g., checking for 'python', 'solution') and do not provide robust sanitization against sophisticated prompt injection attempts.
Audit Metadata