Bring security to every Anthropic product with Speakeasy

Organizations with compliance requirements are struggling to roll out AI tools to their teams. Claude Enterprise rollout has been blocked by security because the controls a compliance program requires were missing. There was no shadow AI detection, so nobody knew which AI tools employees were already using. There was no DLP on the path to the model, so a regulated record could leave in a prompt with nothing to stop it. And there was no chain of record for auditors.

Today we're releasing full security for Anthropic services on the AI Control Plane, so you can roll out Claude to your teams with the security suite your compliance program expects. Claude Chat on desktop, web, and mobile, Claude Design, and Claude Code Web now send every prompt through Speakeasy before the model runs. Speakeasy returns allow or deny from the risk policies and classifiers you already run, and the session lands in Agent Sessions in real time.

Why Claude Enterprise needs an inline gate

A prompt that reaches Claude Chat on the web, Claude Design, or Claude Code Web is already past the endpoint agent, past the browser extension, and past the network filter that never saw the SaaS path. Compliance can export the chat later. It cannot stop the turn that just left.

Anthropic now lets a Claude Enterprise organization name an AI security server and hold every governed request for an allow or deny verdict before inference. Classic DLP products such as Netskope, Palo Alto, Proofpoint, and Zscaler sit on that same hook and reuse the profiles they already run elsewhere. That is a real option if your bar is "extend existing DLP to Claude."

Teams that already point Claude at Netskope or Proofpoint still need a control plane that understands agent sessions and Speakeasy policies. Speakeasy is not another generic DLP destination. It is the judgment layer that already runs your risk policies, PI and prompt-injection classifiers, and session telemetry, so the verdict joins the same chat Compliance will enrich later.

Full security for Anthropic services

One Speakeasy configuration covers Claude Chat on desktop, web, and mobile, Claude Design, and Claude Code Web. You do not install a per-surface agent, and you do not wait for a device to check in.

Every prompt gets a policy verdict before inference. The risk policies and PI and prompt-injection classifiers you already run on the AI Control Plane apply to Claude, so a regulated record is stopped at the point it would leave. That closes the DLP gap on the SaaS path.

Every turn lands in Agent Sessions in real time, and Compliance later enriches the same conversation with authoritative content. The chain of record starts at the prompt instead of at the export, and an auditor gets one conversation with the verdict attached.

How it works

Anthropic Inference hooks is the protocol underneath. A Claude Enterprise organization names Speakeasy as its AI security server, and Anthropic holds every prompt until Speakeasy answers.

Inference hooks are a different mechanism from the hooks in the Claude Agent SDK. Agent SDK hooks are callbacks a developer registers inside their own agent process. They fire on events such as PreToolUse, PostToolUse, and UserPromptSubmit, and they can block a tool call, rewrite its input, or append context before the agent continues. They protect the one agent that developer built, and they run wherever that agent runs. Inference hooks sit on Anthropic's side of the wire. A Claude Enterprise admin names one AI security server for the whole organization, Anthropic calls it over HTTPS for every prompt in Claude Chat, Claude Design, and Claude Code Web.

How to get started

Anthropic Inference hooks on Speakeasy is an ungated beta for Claude Enterprise. You connect once for the organization, and Speakeasy assigns the project and applies its security policies. In the Speakeasy dashboard, open Settings, then AI Integrations, and select Anthropic inference hooks. The panel walks through three steps:

• Copy the webhook URL and paste it into Claude under Organization settings, Data and privacy, Inference hooks. Test the connection and save with Enforce verdicts off.
• Claude reveals a signing secret after you save. Paste it back into Speakeasy so deliveries are authenticated.
• Turn on Enforce verdicts in Claude, set Failure posture to Block and the timeout to 10 seconds, then save. Claude now sends every conversation to Speakeasy for a policy check before inference.

What's next

This release covers real-time session tracking and allow, deny, or review on user prompts and tool calls. Before turning on enforcing mode, know the boundaries of what ships today:

• Speakeasy does not rewrite prompts. There is no rewrite path in the protocol, so Speakeasy does not redact or reshape a prompt. It decides whether the turn proceeds.
• Speakeasy does not block tool responses or assistant responses.
• Image-only attachments are not inspected. Anthropic sends metadata and extracted text, not raw image bytes.
• This path does not distribute MCP servers or plugins, and it does not discover shadow AI on the endpoint. Those remain separate capabilities on the AI Control Plane.
• Coverage is Claude Enterprise only. Personal and Max plans, Bedrock, Vertex, the Claude Platform API, and voice are out of scope.

Inference Hooks are a new integration point, and we will be developing more capabilities as they are released by Anthropic.