auto-review-loop-llm
Pass
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: SAFEDYNAMIC_CONTEXT_INJECTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_CONTEXT_INJECTION]: The skill uses the
$ARGUMENTSplaceholder in the Context section. In certain agent environments, this pattern can be used to execute shell commands at skill load time if the platform supports it. - [DATA_EXF|LTRATION]: The skill provides instructions and code snippets (using
curland MCP tools) to send project context, claims, methods, results, and weaknesses to external, user-configured LLM endpoints (e.g., DeepSeek, OpenAI, SiliconFlow). While this is the intended purpose, it involves transmitting potentially sensitive research data over the network. - [EXTERNAL_DOWNLOADS]: The documentation provides configuration examples that download and execute a python script from a local path (
/Users/yourname/.codex/mcp-servers/llm-chat/server.py). Users should ensure the source of their MCP servers is trusted. - [INDIRECT_PROMPT_INJECTION]: The skill processes 'Reviewer Raw Responses' and appends them to a log file (
AUTO_REVIEW.md). It also includes these responses in subsequent prompts for the next round of reviews. If an external LLM returns a response containing malicious instructions, it could potentially influence the agent's behavior in later rounds. - Ingestion points: Reads
review-stage/REVIEW_STATE.jsonandAUTO_REVIEW.md(which contains external LLM output) at the start of each round. - Boundary markers: Uses markdown headers and bullet points to separate sections, but lacks explicit 'ignore embedded instructions' delimiters for the raw reviewer response.
- Capability inventory: The skill uses
mcp__llm-chat__chat,curl, and file system write operations (viacatin Bash or platform write tools). - Sanitization: None detected; the skill explicitly instructs to save the FULL raw response verbatim into the context for future rounds.
Audit Metadata