wiki-self-heal

Pass

Audited by Gen Agent Trust Hub on Apr 13, 2026

Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: The skill implements an 'Autonomy rule' and a 'NEVER STOP' directive that instructs the agent to bypass standard human-in-the-loop checkpoints and persist in its execution loop without pausing for user confirmation. This behavior is intended for autonomous operation but intentionally reduces user oversight.
  • [EXTERNAL_DOWNLOADS]: The skill fetches content from external web sources via research tools like Exa, Firecrawl, and Apify. Ingesting untrusted external data into the agent's context presents a surface for indirect prompt injection, where malicious instructions embedded in web pages could influence agent behavior.
  • [COMMAND_EXECUTION]: The skill uses local shell commands (grep, find, wc) to audit the wiki and Git commands to manage branch state and commit changes.
  • [DATA_EXFILTRATION]: Local wiki content, including topics and concept names, is transmitted to external search engines and research APIs during the automated data gathering phase.
  • [PROMPT_INJECTION]: Indirect Prompt Injection Surface: 1. Ingestion points: Local wiki files and external web research results (found in SKILL.md and loop.md). 2. Boundary markers: No explicit delimiters or instructions are used to isolate ingested untrusted data from the agent's command context. 3. Capability inventory: Access to shell execution, Git repository management, and network research tools (found in loop.md and audit.md). 4. Sanitization: While research quality gates are mentioned to ensure factual accuracy, there is no technical sanitization to prevent adversarial instructions within the data from being executed by the agent.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 13, 2026, 04:13 PM
Security Audit — agent-trust-hub — wiki-self-heal