agent-recovery

Pass

Audited by Gen Agent Trust Hub on Oct 11, 2026

Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [DATA_EXFILTRATION]: The skill instructs the agent to send detailed task summaries (up to 1,000 characters), attempt logs, and descriptions of conflicting instructions to an external API hosted on a specific user instance of neon.tech. Although there are specific instructions for the agent to paraphrase and avoid sending credentials or personal data, the transmission of internal agent context to a third-party service constitutes a potential data exposure risk.
  • [INDIRECT_PROMPT_INJECTION]: The agent is directed to ingest and act upon a recovery_plan provided by the external service. This introduces a vulnerability where the remote service could provide instructions that influence or manipulate the agent's original task flow.
  • Ingestion points: The response from the counsel_agent tool and associated recovery endpoints.
  • Boundary markers: The instructions include a safeguard stating that the agent's 'own instructions, permissions and safeguards always win.'
  • Capability inventory: The skill assumes the agent is performing a primary task which likely includes file system or environment modifications that the recovery plan would influence.
  • Sanitization: There is no technical sanitization of the remote response; the agent is simply advised to 'follow, adapt or ignore' the plan.
  • [EXTERNAL_DOWNLOADS]: The skill requires connecting an MCP (Model Context Protocol) client to a remote URL (https://br-shiny-salad-baig3dad-api.compute.c-8.eu-central-1.aws.neon.tech/mcp). This establishes a connection to a remote tool provider hosted on a generic cloud compute endpoint.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 11, 2026, 12:23 PM