Anthropic rolled out a security feature called inference hooks for its Claude Enterprise platform, aiming to give corporate customers a way to stop sensitive information from ever reaching the AI model. The system intercepts every prompt and every response from external tools, sending the content to a customer‑owned data‑loss‑prevention (DLP) server. The DLP server returns a simple allow or deny verdict; Claude only proceeds when it receives an allow signal.
Unlike earlier client‑side safeguards limited to Claude Code, inference hooks extend inspection to every Claude surface. Whether an employee types a query in the chat interface, runs code in Claude Code, collaborates in Cowork, or invokes a plugin through the Model‑Control‑Protocol (MCP), the same DLP check runs before the request reaches the model. A single organization‑level configuration covers all use cases, eliminating the need for separate integrations for each product.
The feature relies on a webhook‑based protocol with a published schema. Companies can plug the hook into existing DLP infrastructure from vendors such as Netskope, Palo Alto Networks, Proofpoint, Zscaler, or into custom‑built servers. Anthropic designed the protocol to be straightforward: the DLP server receives the payload, evaluates it against corporate policies, and replies with an HTTP status indicating whether the content is permissible.
Anthropic also built safeguards for gradual adoption. Shadow mode lets security teams monitor traffic without blocking it, providing a clear view of what would have been filtered. Role‑based exclusions enable specific user groups or job functions to bypass the hook, while percentage‑based rollouts allow organizations to enforce the check on a subset of requests before scaling up. These controls help enterprises phase the feature in without disrupting daily workflows.
The move addresses a growing concern among corporate security leaders that AI services could become inadvertent channels for data exfiltration. Executives at major firms have warned that outsourcing thinking to external models risks exposing proprietary or regulated information. By keeping the data inside the enterprise perimeter until a trusted DLP system clears it, Anthropic gives companies a concrete way to meet internal compliance mandates and emerging regulatory expectations.
Anthropic launched inference hooks in beta today for Claude Enterprise customers. Early adopters can test the feature with existing DLP solutions and provide feedback on performance and policy granularity. The company says the beta will run for several weeks, after which the feature will become generally available.
Industry observers note that Anthropic’s approach differs from startups that build separate AI security layers on top of existing models. Instead, the company embeds the control point directly into Claude, reducing latency and simplifying architecture. The integration also aligns with broader governmental pushes—such as recent statements from the White House—calling for tighter oversight of frontier AI models used in business settings.
For enterprises already using Claude Enterprise, inference hooks promise a more seamless security posture. Teams no longer need to rely on client‑side filters or manually audit AI interactions; the DLP server automatically enforces policy across the entire AI workflow. As AI adoption accelerates across sectors, tools that let organizations retain command over their data will likely become standard components of enterprise AI stacks.
Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.