distributed-debugging-debug-trace

Fail

Audited by Gen Agent Trust Hub on Oct 9, 2026

Risk Level: HIGHDYNAMIC_EXECUTIONDATA_EXFILTRATIONCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The remote-debug-server.js file in resources/implementation-playbook.md implements a WebSocket server that accepts arbitrary expressions and executes them within the process using inspector.Session().post('Runtime.evaluate', { expression, ... }). This creates a direct remote code execution (RCE) vector if the service is reachable over the network.
  • [DATA_EXFILTRATION]: Multiple patterns in resources/implementation-playbook.md lead to the exposure of highly sensitive information:
  • The TracingSystem in tracing.js specifically configures instrumentations to capture http.request.body and session.id into spans, which are then exported to a remote Jaeger endpoint.
  • The DebugMiddlewareFactory exposes a manual web route /debug/heap that triggers a heap snapshot. Memory heap snapshots contain environment variables, encryption keys, and raw user data.
  • The ProductionDebugger captures all console logs and returns them to the caller via HTTP headers (X-Debug-Info), creating a data leak path.
  • [COMMAND_EXECUTION]: The Docker configuration provided in the playbook sets NODE_OPTIONS="--inspect=0.0.0.0:9229". Binding the Node.js inspector to all interfaces (0.0.0.0) without authentication is a high-risk configuration that allows remote attackers to attach a debugger and execute arbitrary code.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied requirements via $ARGUMENTS and generates complex diagnostic configurations based on them.
  • Ingestion points: Ingests requirements in SKILL.md to drive the configuration of tracing and debugging workflows.
  • Boundary markers: None identified; the skill does not explicitly warn the agent to ignore instructions embedded within the user's requirements.
  • Capability inventory: The skill generates code capable of file system writes (v8-profiler-next), network operations (JaegerExporter, ElasticsearchTransport), and dynamic code execution (inspector.Session).
  • Sanitization: There is no evidence of sanitization for the inputs used to build the debug dashboard or tracing attributes.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Oct 9, 2026, 02:46 PM