Cisco AI Defense blocks Claude prompts before they run

Anthropic and Cisco have announced an integration that brings AI threat detection directly into Claude Enterprise. The partnership leverages Anthropic's inference hooks—a new capability that lets organizations send each governed prompt to a security service before the model runs—and routes those prompts to Cisco AI Defense, which inspects them for artificial intelligence threats and returns a verdict of allow or deny.

The core problem this solves is a gap in traditional security. As Claude Enterprise users increasingly deploy Claude, Claude Code, and Claude Cowork to act on their behalf—reading documents, calling tools, and running tasks—the prompt itself becomes an attack surface. A jailbreak can attempt to bypass the assistant's guardrails. A prompt injection hidden in a shared file can hijack an agent and turn its own tools against the organization. Traditional data loss prevention tools were built to watch for sensitive data leaving the building; they were never built to see the model being manipulated. That is the exposure Cisco AI Defense targets.

Cisco AI Defense reads intent, not just data. When a user submits a prompt to Claude Enterprise, it is passed to Cisco AI Defense through the inference hook, which inspects the full conversation—including tool calls and their results—against the organization's runtime policy for prompt injection, jailbreaks, tool exploitation, and sensitive data. Every call is cryptographically verified with a one-time signing secret, and Cisco AI Defense confirms authenticity before inspection. A verdict is returned before inference: safe prompts continue; a malicious prompt is blocked, and the model never runs.

The availability timeline shows the integration is live for Claude Enterprise today, with inference hooks currently in beta. The hook fires on each governed prompt before the model runs, which is where enforcement currently happens. As Anthropic extends the hook to responses, the same integration path will enable response-side enforcement. Over time, the enforcement is expected to become a native part of the Cisco AI Defense Inspect API, allowing users to enable it with a few steps in their Claude Enterprise settings. Cisco offers a live playground in its developer portal where users can enter a prompt and watch the real-time inspection and verdict in action.