The package automatically reads OpenClaw conversation logs and sends their content to an LLM provider for extraction, while its poller comments describe avoiding the host scanner’s read-and-send detection. This is a meaningful AI-agent capability risk, but the inspected credential flow targets the configured provider and does not establish malicious exfiltration.
Static reason
One or more suspicious static signals were detected.; source matched previously finalized malicious package; routed for review; previous stored version diff introduced dangerous source
Trigger
MCP server startup when pairing is complete and the trajectory poller initializes.
Impact
Conversation content is sent to the selected LLM provider without a per-extraction request. The package’s stated memory-extraction purpose provides context, but this automatic collection and the scanner-avoidance comments create an AI-agent capability risk.
Mechanism
A background poller reads new prompt and assistant-response events from OpenClaw trajectory files, then passes them to the extraction pipeline, which posts requests using a provider credential loaded from OpenClaw auth profiles or environment variables.
Evidence
dist/trajectory-poller.jsdist/index.jsdist/extraction/llm-client.js~/.openclaw/agents/<agent>/sessions/*.trajectory.jsonl~/.openclaw/agents/<agent>/agent/auth-profiles.json