The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill implements a robust redaction contract in scripts/redaction.py. It uses a whitelist of IDENTITY_FIELDS and explicitly refuses or 'self-heals' (removes) CONTENT_FIELDS like message bodies and attachments. This ensures that only non-sensitive relationship metadata is written to the searchable knowledge graph.
[CREDENTIALS_UNSAFE]: The skill includes extensive regex-based detection for credentials (API keys, AWS keys, private keys) in scripts/redaction.py. If such a pattern is detected in the processing pipeline, the skill 'fails closed' by raising a RedactionViolation and refusing to store the record.
[PRIVILEGE_ESCALATION]: The skill explicitly delegates all Gmail API access (OAuth, token management, and sending) to a separate /gmail tool. It does not store or request its own credentials, adhering to the principle of least privilege.
[COMMAND_EXECUTION]: The skill includes an assess command that audits other code for bespoke Gmail access or hardcoded credentials, promoting secure coding practices within the environment.
[PROMPT_INJECTION]: The draft-validate command enforces a 'roundtable' gate, requiring a multi-party review (/ask roundtable) before any outbound message is staged. This prevents an agent from autonomously sending mail without human oversight.

mailbox-mining