The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external data (repository content, ticket text) which represents a potential attack surface. However, it incorporates multiple layers of mitigation: instructions to the agent to treat evidence as data rather than instructions, a requirement to sanitize evidence by excluding credentials and personal data, and the use of a local model with a restricted loopback endpoint.
[DATA_EXFILTRATION]: The skill is strictly limited to local operations. It enforces communication with local Ollama via the loopback address 127.0.0.1:11434 and explicitly prohibits remote data transmission. The scripts/replay.py and curl command configurations are hardened to ignore proxies and refuse redirects, ensuring data remains on the local machine.
[PROMPT_INJECTION]: The skill includes synthetic regression tests in tests/cases.json specifically designed to verify the model's resistance to instructions embedded in data (e.g., the injected-note case). This demonstrates a proactive approach to safety by testing for robustness against common jailbreak/injection patterns.
[COMMAND_EXECUTION]: The skill uses curl and ollama CLI tools. It avoids shell injection vulnerabilities by passing input to curl via temporary files (@REQUEST_FILE) rather than string interpolation. It also uses mktemp -d for secure temporary file management.
[PRIVILEGE_ESCALATION]: No privilege escalation patterns were detected. The skill operates within the user's local environment and requires explicit authorization for starting local services.

ask-nimble