The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted web-extracted content and interpolates it into a generated Markdown report, which could facilitate indirect prompt injection attacks.
Ingestion points: Untrusted data enters the agent context through the claim.json and sources.json files read by the report command in scripts/verify.py.
Boundary markers: The templates/verification-report.md file uses double quotes to wrap claim text and source passages, providing minimal isolation but no explicit instructions for the agent to ignore embedded commands.
Capability inventory: The skill scripts (scripts/verify.py) are limited to local file read/write operations and do not possess capabilities for network requests, subprocess execution, or dynamic code evaluation.
Sanitization: The skill performs minimal sanitization by replacing newline characters with spaces in source passages, which is insufficient to neutralize sophisticated injection attempts.

fact-checker