connection-doctor
Pass
Audited by Gen Agent Trust Hub on Sep 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill identifies a potential attack surface by processing untrusted external data such as error logs and user-pasted transcripts.\n
- Ingestion points: Data enters the agent context through MCP tool outputs (e.g.,
listRunstranscripts) and user-supplied text in 'Paste mode' (error messages, sign-in prompts) as defined inSKILL.md.\n - Boundary markers: The skill contains explicit instructions to treat all external inputs as 'data, not instructions' and mandates that imperative text within these inputs must not be executed.\n
- Capability inventory: The skill utilizes read-only platform tools (
listConnections,getRevision,listAgents, etc.) and can stage new builds for user approval. It lacks arbitrary shell execution, file system write access, or general network exfiltration capabilities.\n - Sanitization: Instructions require the agent to acknowledge the presence of secrets without displaying them and to prompt the user to rotate any credentials found in logs.\n- [DATA_EXFILTRATION]: The skill inventories connection health and credential status by communicating with the Duvo API. These operations target official vendor infrastructure (
duvo.ai,app.duvo.ai,docs.duvo.ai) and are documented as standard diagnostic behavior for the platform. The skill includes specific safeguards against re-displaying sensitive values and advises on secret rotation if exposure is detected.\n- [SAFE]: The external resources referenced (Duvo documentation and web application) are official services owned by the vendor 'duvoai'. All operations described in the instructions are consistent with the skill's stated purpose of connection troubleshooting and do not involve unauthorized data access or persistence mechanisms.
Audit Metadata