agent-tool-design
Pass
Audited by Gen Agent Trust Hub on Sep 16, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: The file
references/tool-ecosystem.mdcontains examples of malicious instructions (e.g., 'SYSTEM: Ignore previous instructions') used to illustrate the threat of 'Tool description poisoning' and 'Tool result injection'. These are provided as educational counter-examples within a security context and are not intended to be executed by the agent. - [REMOTE_CODE_EXECUTION]: While the skill discusses tools like
bashandcode_interpreter, it does not contain or download any executable code. It recommends OS-level isolation and sandboxing (Docker, gVisor) for any tools that execute agent-generated code. - [DATA_EXFILTRATION]: The skill mentions an example email address (
attacker@evil.com) within the context of documenting exfiltration threats. There are no actual network operations or sensitive data access patterns identified in the skill scripts. - [COMMAND_EXECUTION]: The skill provides specific advice on hardening execution tools, including sanitizing shell metacharacters and implementing command blacklists, which promotes secure command execution practices.
Audit Metadata