llm-council
Warn
Audited by Gen Agent Trust Hub on Oct 31, 2026
Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill contains a 'Codex compatibility note' and a 'Subagent authorization' section with instructions that attempt to define or override agent system behavior. It dictates how skills should be loaded, how tool names should be mapped without user approval, and explicitly claims to self-authorize the use of the
spawn_agenttool once the skill is activated. - [COMMAND_EXECUTION]: The skill instructs the agent to run shell commands using local scripts:
python .claude/scripts/code_graph trace <key-file>andpython .claude/scripts/code_graph connections <file>. These commands utilize variable file paths from the workspace as arguments, which can be risky if those paths are influenced by untrusted data. - [INDIRECT_PROMPT_INJECTION]: The skill has a high vulnerability to indirect prompt injection due to its core workflow. It reads contents from various workspace files (e.g.,
CLAUDE.md,AGENTS.md, and everything indocs/andmemory/) and interpolates them directly into prompts for five parallel advisors, five peer reviewers, and a final synthesis step. - Ingestion points: Files such as
CLAUDE.md,AGENTS.md,docs/project-config.json, and documents within thedocs/project-reference/,docs/specs/, ormemory/directories. - Boundary markers: While it uses
---delimiters in advisor prompts, it lacks comprehensive markers and warnings to ignore embedded instructions across all its multi-stage LLM calls. - Capability inventory: The skill can execute shell commands via Python, read and write files, and spawn multiple sub-agents.
- Sanitization: There is no mention of sanitizing or validating the content of workspace files before they are processed by the various agents.
Audit Metadata