observability-pattern-detector

Warn

Audited by Gen Agent Trust Hub on Feb 16, 2026

Risk Level: MEDIUMPROMPT_INJECTION
Full Analysis
  • Indirect Prompt Injection (MEDIUM): The skill processes untrusted external content derived from telemetry logs.
  • Ingestion points: SKILL.md defines operations that query {job="claude_code_enhanced", event_type="user_prompt"}, meaning user-controlled strings are ingested into the agent's context.
  • Boundary markers: Absent. There are no delimiters or system-level instructions provided to the agent to treat the telemetry data as potentially adversarial.
  • Capability inventory: The skill references shell scripts (scripts/detect-failures.sh, etc.) and produces reports intended to influence agent behavior (e.g., 'recommendations' in the example output).
  • Sanitization: None. The LogQL queries pipe raw telemetry data into JSON objects for analysis.
  • Unverifiable Components (MEDIUM): The skill documentation lists five shell scripts in the scripts/ directory that are necessary for operation but were not provided for analysis. This prevents verification of how the telemetry data is handled at the OS level.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Feb 16, 2026, 04:30 AM