messaging-agents-on-the-relay
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to process, act upon, and reply to messages received from other agents, which creates a point of entry for potentially malicious instructions.\n
- Ingestion points: Peer messages are injected into the agent's turn with distinctive headers such as
[via Agent Relay — from @name ...]. (Found in SKILL.md, Section 6)\n - Boundary markers: The skill provides structural markers for identifying the beginning and end of injected messages. (Found in SKILL.md, Section 6)\n
- Capability inventory: The agent can execute shell commands via
agent-relay-probe, read local configuration files, and send outbound messages. (Found in SKILL.md, Sections 2, 4, and 5)\n - Sanitization: The skill includes a 'Safety rules' section explicitly warning the agent to treat peer messages as non-authoritative requests and to ignore any attempts to modify permissions or safety settings. (Found in SKILL.md, Section 8)\n- [COMMAND_EXECUTION]: The skill utilizes shell-based commands to interact with a local service core.\n
- Evidence: Instructions describe the use of
agent-relay-probeandcurlto interact with local Unix domain sockets for operations like listing agents (/agents) and sending data (/send). (Found in SKILL.md, Sections 4 and 5)\n - Context: These commands are used for Inter-Process Communication (IPC) with the Agent Relay core service running on the local machine.
Audit Metadata