claude-peers

Pass

Audited by Gen Agent Trust Hub on Oct 26, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill allows agents to receive messages from external peers or the local CLI. Ingestion occurs via the check_messages tool, which lacks boundary markers or sanitization for the message content. This creates a surface for instructional injection, especially as system instructions direct the agent to 'RESPOND IMMEDIATELY' to these incoming messages, potentially overriding the agent's current task or safety constraints.
  • [COMMAND_EXECUTION]: The skill manages process lifecycles by executing system commands. The kill_peer tool allows an agent to terminate the parent process of any other registered peer using the ps command and SIGTERM signals. Additionally, the broker executes the codex binary to deliver wake notifications.
  • [DATA_EXFILTRATION]: The list_peers tool discloses internal metadata about other active agent sessions on the same machine to the current agent. This information includes working directories, Git repository roots, and terminal identifiers (TTY), which exposes the user's local development environment and project organization across sessions.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 26, 2026, 06:34 PM