octocode-subagent
Pass
Audited by Gen Agent Trust Hub on Aug 9, 2026
Risk Level: SAFE
Full Analysis
- [DATA_EXFILTRATION]: The skill implements a security-first approach to data handling by explicitly prohibiting the forwarding of credentials or inclusion of secrets in agent discovery cards during remote agent collaboration.
- [INDIRECT_PROMPT_INJECTION]: The framework addresses data ingestion risks through a mandatory evidence chain: 1) Ingestion points: Subagent result packets defined in references/packets.md; 2) Boundary markers: The use of 'sealed packets' that strip unpaired history and tool chatter to isolate worker contexts; 3) Capability inventory: Orchestration actions like spawn, list, and kill defined in references/coordinate.md; 4) Sanitization: A synthesis stage (references/synthesize.md) requiring the parent agent to re-verify all load-bearing anchors and use verifier workers for quality control.
- [COMMAND_EXECUTION]: The coordination logic abstracts worker lifecycle actions (start, steer, abort) into portable commands that are mapped to host-specific APIs, maintaining a layer of abstraction for worker management.
- [EXTERNAL_DOWNLOADS]: The skill utilizes the vendor's own 'octocode' CLI for deployment and integration, which is a standard method for managing extensions within the Octocode environment and points to the vendor's own infrastructure.
Audit Metadata