The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill instructs the agent to search for and read centralized credential files at sensitive paths such as ~/.secrets/keys.txt, ~/.credentials.txt, and ~/Documents/Access.txt.
[CREDENTIALS_UNSAFE]: The skill provides comprehensive regex patterns to harvest authentication tokens for numerous services, including AWS Access Keys, GitHub Tokens, Stripe Secret Keys, and Supabase Service Role Keys.
[DATA_EXFILTRATION]: Extracted credentials are sent to official validation endpoints for services like OpenAI, Anthropic, and Reddit using curl. This process involves transmitting sensitive master keys from a local store to external servers.
[COMMAND_EXECUTION]: The skill utilizes shell commands (curl) to perform network requests and cat to write environment variables, interpolating extracted credentials directly into the execution string.
[PROMPT_INJECTION]: The skill processes untrusted data from local files and incorporates it into agent instructions and shell commands without specific boundary markers or sanitization, creating a surface for indirect prompt injection. Findings for this category involve: (1) Ingestion point: SKILL.md via parse_credentials_file; (2) Boundary markers: Absent; (3) Capability inventory: Subprocess calls via curl, file-write via cat in SKILL.md; (4) Sanitization: Absent.

credentials