The Agent Skills Directory

[SAFE]: The skill is a standard operational tool developed by PostHog. It relies on internal platform tools (alert-simulate, execute-sql, scout-scratchpad) to perform metric monitoring and maintain state between runs. All observed behaviors align with its stated purpose.
[COMMAND_EXECUTION]: The skill uses Python 3 for advanced statistical analysis, specifically Kolmogorov-Smirnov tests for distribution shifts. This involves executing a local script (scripts/ks2.py) or inline heredocs. These operations occur within the agent's sandbox, processing numeric data retrieved from PostHog, and do not pose a risk of system-level compromise.
[DATA_EXFILTRATION]: Data access is limited to the PostHog project environment via tools like insight-query and execute-sql. Reports are emitted to the PostHog report channel (scout-emit-report). No network calls to third-party domains or unauthorized data transfers were identified.
[PROMPT_INJECTION]: The instructions are professional and focused on the technical implementation of anomaly detection. There are no attempts to bypass safety filters or manipulate the agent's underlying behavior.
[PROMPT_INJECTION]: The skill is evaluated for indirect prompt injection risks (Category 8) as it processes external event data. The risk is assessed as SAFE due to the intended use and specific data handling.
Ingestion points: Event data and insight results are ingested via execute-sql and insight-query as described in SKILL.md and references/anomaly-methods.md.
Boundary markers: While explicit delimiters are not mandated for raw SQL results, the instructions focus on extracting numeric values and histograms.
Capability inventory: The skill possesses python3 execution capabilities and the ability to file reports via scout-emit-report.
Sanitization: Ingested data is parsed as JSON and processed through numeric mathematical functions in scripts/ks2.py, which naturally mitigates non-numeric injection payloads.

signals-scout-anomaly-detection