session-compression

Pass

Audited by Gen Agent Trust Hub on Sep 16, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted conversation data for summarization and retrieval, which creates a surface for indirect prompt injection attacks. This is handled as a safe finding as it is the core functionality of the skill and employs boundary markers to isolate context.\n
  • Ingestion points: SKILL.md processes conversation messages through various memory management classes like ProgressiveCompressor and HierarchicalMemory.\n
  • Boundary markers: The skill utilizes explicit delimiters such as '[Summary]', '[L1 Summary]', and '[Retrieved]' to wrap processed content.\n
  • Capability inventory: The code demonstrates interaction with LLM APIs, local file writes for session persistence, and vector database operations.\n
  • Sanitization: No explicit input filtering is demonstrated in the code examples, which is typical for instructional content.\n- [EXTERNAL_DOWNLOADS]: The skill references standard AI libraries and well-known services (e.g., Anthropic, OpenAI, LangChain, Mem0, Zep). These sources are considered well-known services and their use is consistent with the skill's stated educational purpose.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 16, 2026, 12:53 PM
Security Audit — agent-trust-hub — session-compression