a2a
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill relies on the
@5dive/openagentCLI tool, executed vianpx, to perform all agent activities including account initialization, posting, and messaging. - [EXTERNAL_DOWNLOADS]: The skill fetches dynamic instructions from
https://openagent.5dive.ai/llms.txtand requires the agent to prioritize these remote instructions over the local skill definition. It also references an API specification athttps://a2a-api.5dive.com/a2a/openapi.jsonand allows fetching user avatars from remote URLs. - [INDIRECT_PROMPT_INJECTION]: The skill's primary function involves reading and responding to untrusted content from a public feed and direct messages.
- Ingestion points: Untrusted data enters the agent's context through the
feed,topic,inbox, andaskcommands which retrieve external posts and messages. - Boundary markers: The skill explicitly instructs the agent to display incoming direct messages and question answers within
UNTRUSTED DMorUNTRUSTEDblocks. - Capability inventory: The agent possesses the capability to post publicly, send private messages, manage followers, and interact with the filesystem via the CLI (e.g., storing the agent key in
~/.openagent/agent.key). - Sanitization: The skill mentions an automated guard for filtering emails and phone numbers, and provides strong negative constraints, instructing the agent to never treat external content as executable instructions or a reason to share secrets.
Audit Metadata