research-archival

Fail

Audited by Gen Agent Trust Hub on Oct 8, 2026

Risk Level: HIGHCREDENTIALS_UNSAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [CREDENTIALS_UNSAFE]: The skill attempts to resolve GitHub authentication by scanning sensitive local file paths, specifically ~/.claude/.secrets/gh-token-*, and by executing ~/.claude/tools/bin/gh-token-for-repo to retrieve tokens.
  • [COMMAND_EXECUTION]: The identity verification and scraping workflows utilize complex Bash heredoc blocks that execute various CLI utilities including gh, curl, git, and jq while manipulating environment variables and system configuration directories.
  • [DYNAMIC_EXECUTION]: The skill includes a 'Self-Evolving Skill' directive that instructs the agent to use the Edit or Write tools to modify the SKILL.md file itself to address instruction drift or workarounds, which constitutes self-modifying code.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from untrusted external URLs (via Firecrawl and Jina Reader) and process this content for storage in local markdown files and GitHub Issue bodies, creating a surface for indirect injection.
  • Ingestion points: API responses from api.firecrawl.dev and r.jina.ai containing scraped markdown content from arbitrary user-provided URLs.
  • Boundary markers: The skill employs a specific artifact separation strategy where raw scraped transcripts are stored in separate .raw-scrape.txt files and the main archival .md file contains a 'provenance warning'.
  • Capability inventory: The skill possesses the capability to write to the filesystem (Write tool) and interact with GitHub APIs (gh CLI) to create and comment on issues.
  • Sanitization: The workflow uses jq to structure remote JSON data and implements sha256 hashing to ensure the integrity of the raw archived artifacts.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Oct 8, 2026, 02:13 AM