The Agent Skills Directory

[DATA_EXFILTRATION]: The skill incorporates a mechanical 'leak guard' (lib/sanitizer.py) that acts as a Data Loss Prevention (DLP) layer. It scans proposed issue content for sensitive patterns like file paths, import statements, and fenced code blocks, blocking the operation if they are found. This effectively prevents the leakage of private code or configuration into public issue trackers.
[COMMAND_EXECUTION]: External commands are invoked via subprocess.run using argument lists instead of shell strings. This design prevents command injection attacks, as user-provided data (such as issue titles) is never interpreted by a shell.
[CREDENTIALS_UNSAFE]: GitHub tokens are handled securely through a targeted injection mechanism in lib/cli.py. A sensitive access token (ISSUES_TOKEN) is only provided to the GitHub CLI when it is specifically targeting the trusted dividedby/skills repository, minimizing the surface area for credential exposure.
[PROMPT_INJECTION]: Risks associated with processing external research data (Indirect Prompt Injection) are mitigated by the fact that the agent's proposed improvements are subjected to mechanical sanitization and a budgeted gating process before being published.
[EXTERNAL_DOWNLOADS]: The skill clones a public research mirror and its own source code from the author's verified repository. These are documented, purpose-aligned downloads from trusted vendor infrastructure.

apply-agent-research