ChatGPT Triggered 410 Million Data Leak Alerts Last Year. Anthropic Just Built a Checkpoint For That.

On August 5, Anthropic quietly shipped one of the more consequential enterprise AI features of the year. It's not a smarter model. It's a checkpoint. Inference hooks routes every prompt sent to Claude Enterprise through the customer's own security server first, for an allow or deny verdict, before the model generates a single token. To understand why that's worth paying attention to, look at the number Anthropic didn't put in its own announcement: 410 million.

The exposure this actually addresses

That's how many data-loss-prevention policy violations Zscaler's ThreatLabz researchers tied to ChatGPT alone in 2025, out of an analysis covering roughly 9,000 organizations and nearly a trillion AI and ML transactions. AI activity across that dataset was up 83% year over year. The violations Zscaler tracked weren't exotic attacks. They were employees pasting source code, medical records, and Social Security numbers into a prompt box with nothing standing between them and the send button. Zscaler's own report frames the problem plainly: AI tools have become the primary channel for accidental corporate data exposure, and most organizations still can't see it happening.

What inference hooks actually does

Every inference request now routes through a signed WebSocket connection to a security server the customer controls. Before Claude starts generating a response, it sends the prompt and its surrounding context to that server, which returns a binary verdict: allow or deny. The server can't rewrite or redact a prompt, only block it. Coverage spans chat, Claude Code, and Claude Cowork, and it applies to tool calls made through MCP connectors, skills, and plugins, all from a single organization-level configuration.

The protocol is open and webhook-based, and Anthropic named four security platforms customers can point it at out of the box: Netskope, Palo Alto Networks, Proofpoint, and Zscaler, the same company whose research produced the 410 million figure above. Companies can also route to an in-house server. Rollout options include shadow mode, which logs without blocking, plus role-based exclusions and percentage-based rollouts, so security teams don't have to flip the switch for the whole org on day one.

Andrew Grimmett, VP of Information Security at Bandwidth, one of the companies Anthropic quotes in the announcement, described the value in operator terms: "Inference hooks add a checkpoint to inspect what's flowing to Claude in real time, before the model ever sees it. This lets us safely move faster on AI without giving up control."

Why this was never really a model problem

It's tempting to read AI data leakage as a reason to distrust the model. That's the wrong read. The 410 million violations Zscaler tracked didn't happen because ChatGPT is unusually unsafe. They happened because almost nothing in the enterprise stack was built to inspect what an employee types into a chat box before it leaves the network. Claude, ChatGPT, or any other frontier model would produce the same exposure without a checkpoint in front of it. <b>Anthropic isn't claiming Claude is safer. It's admitting the gap was never going to close at the model layer, and building the layer that actually closes it.</b>

The catch most coverage is skipping

Inference hooks routes to a security server the customer already runs. If your organization doesn't have a compatible DLP platform in place, or the engineering capacity to stand up a webhook-based one in-house, there is nothing for inference hooks to point at. This is a control that mature security stacks unlock, not a feature that arrives pre-installed and ready for everyone. Companies without that infrastructure now face a build decision before they can use this at all, and that decision sits upstream of the AI rollout, not downstream of a breach.

What to do with this if you're evaluating Claude Enterprise

• Ask any AI vendor whether their enterprise plan inspects prompts before the model sees them, not just outputs after the fact. Inbound inspection is the harder, less common capability.
• If you're already on Claude Enterprise, check whether inference hooks maps to a security platform you run today. That's the fastest path to turning it on, likely in shadow mode first.
• If you don't have a DLP platform in the loop yet, treat that as a prerequisite decision for your AI rollout, not an afterthought you get to once something goes wrong.

Where TrueHorizon fits

This is exactly the kind of decision we help enterprise teams make correctly the first time. We wire governance, access control, and data inspection into agentic AI deployments from day one, rather than retrofitting them after a leak forces the conversation. <b>That's not a checklist we're learning on your project. It's the expertise we bring to it.</b> Anthropic just proved the industry agrees the checkpoint matters more than the model. We've been building that checkpoint into client deployments already.

If you're rolling out Claude, ChatGPT, or any agentic AI platform and you don't yet know what happens to a sensitive prompt before it reaches the model, take our AI readiness assessment before your next rollout, not after your first incident.