llm-council-failure-modes
Pass
Audited by Gen Agent Trust Hub on Sep 7, 2026
Risk Level: SAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill analyzes the attack surface where untrusted data (like tool results) can introduce prompt injections that propagate between agents. It provides technical defenses including provenance tracking (genealogy-graphs) and treating inter-agent messages as untrusted inputs.
- [DATA_EXFILTRATION]: To prevent exfiltration, the skill recommends a 'least-privilege' model where deliberation agents are denied direct access to network or file tools, restricting tool execution to a separate authorization layer.
- [COMMAND_EXECUTION]: The skill warns against granting tool access per-agent, suggesting instead that permissions should be scoped to individual tool calls to prevent a single compromised agent from inheriting a full suite of dangerous capabilities.
- [SAFE]: No obfuscation, hidden code, or persistence mechanisms were found. The skill is entirely advisory, utilizing markdown-based references to academic security concepts to guide safer multi-agent orchestration.
Audit Metadata