The Agent Skills Directory

[SAFE]: The skill provides tools for monitoring inference-API usage and detecting patterns associated with model stealing (AML.T0024.002) and membership inference (AML.T0024.000). The logic implemented in scripts/agent.py performs local analysis of audit logs and uses the Adversarial Robustness Toolbox (ART) for red-teaming.
[EXTERNAL_DOWNLOADS]: The skill recommends installing standard, well-known libraries from the Python Package Index (PyPI), specifically adversarial-robustness-toolbox, scikit-learn, and numpy. These are recognized tools in the AI security community.
[COMMAND_EXECUTION]: The provided Python script uses standard argument parsing and performs local computations (hashing, statistical scoring) without spawning shell processes or executing external commands via os.system or subprocess.run.
[DATA_EXFILTRATION]: No network exfiltration patterns were detected. The script logs inference data locally to inference_audit.jsonl and does not transmit sensitive information to external domains.
[PROMPT_INJECTION]: There are no instructions directed at an LLM that would override safety guidelines or bypass behavioral constraints.

detecting-model-extraction-attacks