self-improving-agent-extract

Pass

Audited by Gen Agent Trust Hub on Jul 18, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses shell commands including grep, sed, and pwd to programmatically access and search Claude's internal project memory storage located at $HOME/.claude/projects/. This behavior is central to the skill's purpose of extracting patterns from previous work but involves reading sensitive internal metadata files.
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It ingest untrusted data from two sources:
  • User-provided pattern descriptions
  • Historical project memory entries found via grep searches
  • Capability Inventory: The skill creates new files, spawns sub-agents (skill-extractor), and interacts with the clawhub publishing tool.
  • Boundary Markers: There are no explicit instructions or delimiters defined to prevent the agent from obeying malicious instructions embedded within the memory entries it is processing.
  • Sanitization: No sanitization or validation of the source text is performed before it is used to generate the new SKILL.md files.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 18, 2026, 04:18 PM
Security Audit — agent-trust-hub — self-improving-agent-extract