Anthropic Compliance API Exposes Local Claude Code Agents Running on Developer Endpoints with Inherited Credentials

Anthropic released new Compliance API endpoints in August 2026 that capture local session transcripts from Claude Code installations. These endpoints record file reads, shell commands, and MCP tool invocations on developer machines. The API provides visibility into harness activity that the LLM itself does not retain, yet it stops short of mapping those actions to the actual credentials or permissions present on the host. Token Security data shows local agents comprise 68.6 percent of discovered AI agents and inherit employee network position by design.

Endpoint telemetry from EDR tools reveals processes and configurations that cloud logs miss, but it supplies evidence rather than governance. A Cloud Security Alliance survey of 418 professionals found 82 percent had uncovered unknown agents in the prior year despite 68 percent claiming high visibility. Anthropic's managed settings JSON records offer a policy baseline, yet they cannot prevent the harness from performing destructive actions when credentials are already present on the machine.

The architecture splits the LLM brain in Anthropic's cloud from the execution harness on the endpoint. This design shifts responsibility onto local controls because classic SaaS dashboards cannot observe runtime behavior or third-party connections made through stolen or cached credentials. Infostealer malware already targets these same developer endpoints, amplifying the gap between logged activity and verified intent.

Organizations will need to correlate Anthropic transcripts with endpoint detection and identity governance layers within the next two quarters to close the authorization blind spot. Without that fusion, activity logs will continue to document actions without confirming legitimacy.