self-improvement
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONREMOTE_CODE_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface by instructing the agent to process external data (errors and corrections) into persistent learning logs.\n
- Ingestion points: Tool outputs captured via environment variables in
scripts/error-detector.shand user corrections/feedback parsed according to instructions inSKILL.md.\n - Boundary markers: Absent. The reminders injected into the session (e.g.,
<error-detected>and<self-improvement-reminder>) do not include instructions for the agent to ignore potentially malicious content embedded within the data being evaluated.\n - Capability inventory: The skill possesses file-writing capabilities to the
.learnings/directory and executes local shell scripts (scripts/*.sh) as part of its hook system.\n - Sanitization:
scripts/extract-skill.shperforms basic relative path checks to prevent directory traversal, but the content logged from tool outputs or user feedback is not sanitized before being written to markdown files.\n- [COMMAND_EXECUTION]: The skill relies on shell scripts for core functionality, including the activator hook and the error detection utility.\n - Evidence:
scripts/activator.sh,scripts/error-detector.sh, andscripts/extract-skill.share intended for local execution.\n- [REMOTE_CODE_EXECUTION]: Installation instructions direct the agent to fetch code from external, non-verified sources.\n - Evidence:
SKILL.mdrecommendsclawdhub install self-improving-agentandgit clone https://github.com/peterskoett/self-improving-agent.git.
Audit Metadata