caveman-manage
Pass
Audited by Gen Agent Trust Hub on Aug 11, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: No malicious patterns, obfuscation, or unauthorized data exfiltration techniques were detected.
- [COMMAND_EXECUTION]: The skill utilizes a vendor-specific CLI tool (
caveman) and MCP tools (caveman_experiment_get) exclusively for read-only operations such as listing experiments and fetching results. No destructive or mutative commands are present. - [PROMPT_INJECTION]: The instructions include a set of "Non-negotiable gates" which function as internal safety guardrails. These instructions reinforce the agent's read-only scope and prevent it from making autonomous decisions on sensitive production controls, which is a defensive best practice.
- [DATA_EXFILTRATION]: The skill demonstrates security awareness by instructing the agent not to supply organization IDs, instead relying on the logged-in identity and server-side Role-Based Access Control (RBAC) to define scope.
- [INDIRECT_PROMPT_INJECTION]: While the skill ingests external data (experiment evidence and results), the risk is mitigated by the explicit block on executable mutations. The agent is restricted to generating recommendations rather than executing code based on the ingested content.
Audit Metadata