autonomous-agent-patterns
Pass
Audited by Gen Agent Trust Hub on Jul 10, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [SAFE]: The skill is a documentation-only resource consisting of markdown text and Python code templates. It does not include executable shell commands or autonomous network operations.
- [COMMAND_EXECUTION]: Sections 3.3 and 5.2 describe implementation patterns for executing shell commands via subprocess. The provided examples include security best practices such as command whitelisting, home directory isolation, and path validation.
- [PROMPT_INJECTION]: The skill patterns the ingestion of external data from files and URLs (Section 5.1), creating a potential surface for indirect prompt injection. Mandatory Evidence Chain: (1) Ingestion points: ContextManager.add_url (SKILL.md) fetches content from external URLs; (2) Boundary markers: Markdown headers and code blocks in format_for_prompt (SKILL.md) serve as delimiters; (3) Capability inventory: ReadFileTool, EditFileTool, run_command, and MCPAgent.create_tool (SKILL.md); (4) Sanitization: SandboxedExecution (SKILL.md) and ApprovalManager (SKILL.md) provide patterns for command validation and human-in-the-loop review.
- [DATA_EXFILTRATION]: Section 2.1 and 5.1 provide implementation templates for reading local files and fetching remote URLs. These are presented as educational examples for agent capabilities and do not contain instructions for unauthorized data transfer.
Audit Metadata