The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted corpora such as social media archives and personal messages which could contain malicious instructions. The skill includes explicit security boundaries in references/source-packets.md to treat this data as untrusted and avoid executing embedded commands. Ingestion points include scripts/prepare-x-archive.ts and scripts/validate-source-packet.ts. Capability inventory includes file reading and synthesis. Sanitization instructions are present but rely on LLM adherence to boundaries.
[COMMAND_EXECUTION]: The skill invokes local utility scripts using bun for archive processing and schema validation. These scripts (prepare-x-archive.ts, validate-source-packet.ts) are part of the skill's trusted distribution and focus on local parsing tasks.
[DATA_EXFILTRATION]: The core function of the skill involves processing sensitive personal data from a user-provided corpus. The instructions and scripts emphasize local-only processing and use restrictive file permissions (0o600) for output files to prevent unauthorized exposure of the generated model and source packets.

ensoul