self-managed-context
Pass
Audited by Gen Agent Trust Hub on Oct 1, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill references and describes functionality from the official Facebook Research repository for Context Language Models. These references are for academic implementation purposes.
- [INDIRECT_PROMPT_INJECTION]: The context-as-a-file architecture creates a surface where the model reads its own previous outputs as potential instructions. The documentation acknowledges this 'self-injection' risk and implements mitigations including pinned system prefixes and role folding.
- Ingestion points: A local mirror file at /tmp/.live_ctx/LIVE_CTX_MAIN.txt.
- Boundary markers: Explicit headers like [[CTX_TURN n role=...]].
- Capability inventory: Shell access for context editing tools.
- Sanitization: Role folding (converting model-written roles to 'user') and pinned system messages.
- [DYNAMIC_EXECUTION]: The skill provides examples of the agent using Python scripts to perform regex-based edits on the mirror context file.
Audit Metadata