Cisco AI Defense now integrates with Anthropic’s Claude Enterprise via inference hooks, which send each governed prompt to the security service before inference begins. Claude, Claude Code, and Claude Cowork act on behalf of users by reading documents, calling tools, and running tasks, making the prompt an attack surface. A jailbreak can bypass guardrails, and a prompt injection hidden in a shared file can hijack an agent and turn its tools against the user. Traditional data loss prevention tools watch for sensitive data leaving but cannot see model manipulation. When configured, the inference hook calls Cisco AI Defense, which inspects the prompt for AI threats and returns a verdict: allow or deny. Safe prompts go through; prompts carrying injection or jailbreak are blocked, and the model never runs. Cisco AI Defense reads intent, catching prompt injection and jailbreaks targeting the model, its tools, and agents, and evaluates agent activity in the conversation transcript. When Claude Code or Cowork calls a tool, including over the Model Context Protocol (MCP), the transcript includes the tool call and result, so poisoned content is caught before the next inference. Cisco AI Defense runs privacy and manipulation guardrails on the transcript, flagging sensitive data plus injection and jailbreak attempts. The integration is one runtime enforcement point in a broader platform that discovers and inventories AI assets, assesses risk through validation, red-teaming, and supply chain and model scanning, and protects at runtime. It is part of Cisco Cloud Control.