The Agent Skills Directory

[DYNAMIC_EXECUTION]: The IsolationForestDetector class in references/EXAMPLES.md employs joblib.load() to restore trained models from disk. Because joblib uses pickle internally, it is susceptible to unsafe deserialization. An attacker who can provide or replace the .pkl model file could achieve arbitrary code execution on the host environment when the skill loads the model.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources, including Prometheus API endpoints and local CSV files, via the MetricsDataLoader class. 1. Ingestion points: MetricsDataLoader.load_from_prometheus and MetricsDataLoader.load_from_csv in references/EXAMPLES.md. 2. Boundary markers: None identified for metric data ingestion. 3. Capability inventory: Uses Bash for environment setup and requests for network operations. 4. Sanitization: Missing sanitization for external content beyond basic numeric cleaning (e.g., np.nan_to_num). While the data is used for mathematical models, the lack of boundaries for ingested external content remains an architectural risk.
[DATA_EXFILTRATION]: The IntelligentAlerter class uses the requests library to transmit incident details to Slack and PagerDuty. While functional, this represents a pathway for data to leave the local environment. Users should verify that these destinations are trusted and that credentials (webhooks/API keys) are managed as secrets.

detect-anomalies-aiops