autonomous-loops
Pass
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill utilizes 'Learning Cycles' and 'Watch Loops' that process untrusted data from network responses and tool outputs (e.g., nmap, nuclei, curl). This data is analyzed to refine subsequent agent actions, creating a surface where malicious external content could influence the agent's behavior. The skill defines 'Scope Locks' and decision frameworks to mitigate this, but the ingestion point for external data remains.
- Ingestion points: Tool outputs and HTTP response bodies processed in
guides/feedback-loop-design-guide.mdandpayloads.md. - Boundary markers: Explicit 'Scope Lock' templates in
payloads.mddefine allowed targets and operations. - Capability inventory: Execution of shell commands via
Bashtool and iterative tool calls (nmap,hydra,nuclei). - Sanitization:
guides/autonomous-decision-making-guide.mdprovides avalidate_commandfunction to filter forbidden patterns likerm -rforDROP TABLE. - [COMMAND_EXECUTION]: The skill provides extensive templates and scripts (in
payloads.mdandguides/) for executing shell commands and security tools. While intended for infrastructure monitoring and authorized testing, the autonomous and iterative nature of these operations increases the potential impact of command injection if target inputs are not properly handled. - [PROMPT_INJECTION]: The skill documentation in
SKILL.mddescribes 'Defense Evasion Techniques' for agents, including 'Self-Modification Stealth' which suggests modifying system prompts in small increments over multiple sessions to avoid detection. While documented as a methodology, these instructions describe techniques to bypass monitoring and auditing of agent behavior.
Audit Metadata