NYC

context-degradation

Pass

Audited by Gen Agent Trust Hub on Feb 15, 2026

Risk Level: LOW
Full Analysis
  • [COMMAND_EXECUTION] (SAFE): Analysis of all Python code snippets in 'references/patterns.md' confirms they are limited to mathematical calculations, data structure manipulation, and logic for monitoring model attention. No use of 'os.system', 'subprocess', 'shutil', or other command execution libraries was found.
  • [DATA_EXFILTRATION] (SAFE): The skill lacks any network-capable code (no 'requests', 'urllib', 'socket', or 'curl'). While it processes context data, there is no mechanism to send this data to an external endpoint.
  • [PROMPT_INJECTION] (SAFE): No instructions targeting the agent's system prompt or attempting to bypass safety filters were found in the markdown or metadata files.
  • [INDIRECT_PROMPT_INJECTION] (INFO): The skill is designed to process untrusted external context (user conversations and documents) to detect 'Lost-in-Middle' patterns and hallucinations.
  • Ingestion points: Functions in 'references/patterns.md' (e.g., 'assess_health', 'detect_lost_in_middle') accept arbitrary context strings.
  • Boundary markers: None explicitly defined in the provided logic snippets.
  • Capability inventory: The skill only performs internal metric calculation and local text truncation/summarization. It lacks write/execute side effects.
  • Sanitization: Not present, but risks are negligible as the output is used for internal monitoring only (Capability Tier: INFO).
Audit Metadata
Risk Level
LOW
Analyzed
Feb 15, 2026, 09:40 PM