evaluate
Pass
Audited by Gen Agent Trust Hub on Oct 17, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONDYNAMIC_EXECUTIONOBFUSCATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches and executes the ouroboros-ai package via the uvx tool from a public package registry as part of its fallback mechanism.
- [REMOTE_CODE_EXECUTION]: Downloads and executes external code from a public registry when the primary MCP server is unavailable.
- [DYNAMIC_EXECUTION]: Utilizes a Python one-liner script executed via uvx to access internal package resources at runtime.
- [OBFUSCATION]: Employs Base64-encoded state payloads embedded in HTML comments to transmit session data between different agent instances or turns.
- [INDIRECT_PROMPT_INJECTION]: Processes untrusted code artifacts and external prompt templates which may contain malicious instructions that could influence subagent behavior.
- Ingestion points: Reads user-specified artifact files and loads evaluator.md from external sources.
- Boundary markers: Lacks specific delimiters or instructions to ignore embedded commands within the ingested content.
- Capability inventory: Possesses access to run_shell, inspect_code, and orchestrate_subagents for acting on instructions.
- Sanitization: No explicit sanitization or filtering of the artifact content is performed before it is analyzed by the semantic review or consensus stages.
Audit Metadata