Anthropic's New Compliance API for Claude Code: What It Shows, What It Misses

Key points

• Anthropic has released new Compliance API endpoints for Claude Code, giving security teams their first direct feed of agent activity on developer machines.
• Claude Code can read files, execute shell commands, and call outside tools using whatever credentials the developer already has.
• Activity logs show what an AI agent did, not whether the access itself was legitimate under company policy.
• Governance frameworks such as the NIST AI Risk Management Framework and the EU AI Act push organisations to document and control automated decision systems, including coding agents.
• Security teams are being told to treat AI coding assistants as a new class of identity, not as a feature bolted onto an existing developer account.

Anthropic has quietly opened a door security teams have been asking about for months. Its new Compliance API endpoints for Claude Code, the company's AI coding assistant, let enterprises pull activity records showing what the agent did on a developer's machine: files it read, shell commands it ran, outside tools it called through the Model Context Protocol, a standard way for AI assistants to plug into other software.

That visibility matters. It's also not enough on its own.

What does the Compliance API actually show?

The endpoints give security and compliance teams a record of Claude Code's actions inside an organisation. Until now, most of that activity sat on individual laptops with no central log. The new feed changes that, at least for customers on Anthropic's enterprise plans.

Think of it as a flight recorder. It captures what happened but tells you nothing about whether the flight should have taken off.

Claude Code operates with the same credentials as the developer running it. If a developer has access to a production database, so does the agent on their machine. The Compliance API will show you the query. It won't tell you whether the agent should have been trusted to run it.

Why is this a policy problem, not just a logging problem?

AI agents are starting to look like a new kind of user, and most identity governance rules were written before they existed. Regulators are catching up. The NIST AI Risk Management Framework, published in January 2023, asks organisations to map and manage risks from AI systems that act on behalf of humans. The EU AI Act, which entered into force on 1 August 2024, imposes staged obligations on providers and deployers of AI systems, with general-purpose model rules applying from 2 August 2025 and high-risk system obligations phasing in through 2027.

Neither framework was written with coding agents specifically in mind. Both apply.

Accountability is where things get uncomfortable. When an AI agent running under a developer's identity deletes a table or exposes a secret, most companies today treat that as the developer's problem. That answer won't hold for long. Our 11 August story found that almost half of companies already say AI governance problems are slowing their AI work down, and coding agents sitting outside any formal identity boundary are exactly the kind of gap those surveys are measuring.

What should security teams do now?

Start by treating the coding assistant as its own identity, even when it borrows a human one. Pull the Compliance API feed into the same monitoring pipeline as other privileged activity, and write policies that say what an agent may or may not do without a human in the loop.

The harder work is upstream. If a developer account has standing access to production, the agent inherits that access the moment it launches. Just-in-time credentials and scoped service identities for AI tooling both help. So does a plain-English written policy telling engineers when it's acceptable to point Claude Code at customer data.

Logs answer what happened. Governance answers what should have been allowed. Anthropic's given enterprises the first. The second is still on the customer.