The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted event data from PostHog (such as event names and properties) which could theoretically contain malicious instructions.
Ingestion points: Data enters the context via execute-sql and insight-query as described in SKILL.md and references/source-liveness.md.
Boundary markers: The instructions explicitly warn against direct interpolation of untrusted strings into queries.
Capability inventory: The skill can execute SQL, run local Python/Bash scripts, and author reports within the PostHog platform.
Sanitization: The skill employs cityHash64 to hash properties like $host and $lib before using them in logic, effectively preventing the agent from processing the raw content of untrusted strings as instructions.
[DYNAMIC_EXECUTION]: The skill generates and runs local Python code for statistical calculations.
Evidence: references/anomaly-methods.md contains a template for an inline Python heredoc to perform Kolmogorov-Smirnov tests. It also utilizes a bundled script scripts/ks2.py. This behavior is restricted to local data processing and does not involve loading external or untrusted code.
[COMMAND_EXECUTION]: The skill uses Bash and python3 to facilitate its core analytical functions. These operations are performed within the agent's restricted sandbox to handle data that has been retrieved through official platform tools.

signals-scout-anomaly-detection