context-compression
Pass
Audited by Gen Agent Trust Hub on Feb 17, 2026
Risk Level: SAFE
Full Analysis
- Prompt Injection (SAFE): No instructions found that attempt to bypass safety guidelines, extract system prompts, or override agent behavior instructions.
- Data Exposure & Exfiltration (SAFE): No sensitive file access or unauthorized network operations. Use of environment variables is limited to retrieving context window configuration limits.
- Remote Code Execution (SAFE): No remote scripts are downloaded or executed. The skill does not install external dependencies or execute code on untrusted data.
- Dynamic Execution (SAFE): Code snippets demonstrate local data processing and template generation without unsafe use of eval() or dynamic loading from untrusted paths.
- Indirect Prompt Injection (SAFE): The skill processes untrusted conversation history for summarization. 1. Ingestion points: Processes message history in references/priority-management.md and assets/compression-probes-template.md. 2. Boundary markers: Implements structured headers (e.g., '## Session Intent') in templates to logically separate content. 3. Capability inventory: State synchronization and summarization; no shell or system capabilities are exposed. 4. Sanitization: No explicit sanitization of history is performed, but the structured anchoring method and functional probe validation provide oversight.
Audit Metadata