flow-deep
Fail
Audited by Gen Agent Trust Hub on Sep 28, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill documents and utilizes a remote code execution pattern in
references/capability-registry.mdfor installing theprime-agenttool viacurl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh. This executes an external script directly in the shell without prior integrity verification. - [DATA_EXFILTRATION]: The skill accesses highly sensitive local files to harvest configuration data and session history.
SKILL.md(Stage 0) andreferences/capability-registry.md(C34) instruct the agent to inspect~/.claude/settings.jsonand~/.prime/agent/auth.jsonto extract API keys (specifically thezai api_key).- The script
scripts/check_context.pyprogrammatically locates and reads Claude Code's internal project transcripts in~/.claude/projects/, which contain the full history of user interactions. - [DYNAMIC_EXECUTION]: The skill dynamically assembles and injects executable prompt instructions into sub-agents based on runtime task metadata (
agent_hint.type), as defined inreferences/skill-routing.md. This includes instructions that dictate the sub-agent's use of specific tools and workflows. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process arbitrary external data from codebases and the web, creating an attack surface for indirect injection.
- Ingestion points: Untrusted data enters the context via
Glob,Grep,Read,WebFetch, andWebSearchtools. - Boundary markers: While the skill uses a "Goal Contract" to define objectives, it lacks strict technical delimiters or escaping mechanisms to prevent processed data from influencing agent behavior.
- Capability inventory: The skill possesses broad capabilities, including file modification, command execution, and the ability to dispatch further agents.
- Sanitization: There is no evidence of input sanitization or validation of external content before it is interpolated into sub-agent prompts.
Recommendations
- AI detected serious security threats
Audit Metadata