Cisco enhances enterprise AI protection by linking its AI Defence system to Anthropic’s Claude Enterprise, aiming to detect and prevent prompt injection and jailbreak attempts before they impact AI workflows.
Cisco is tightening its grip on enterprise AI security as Anthropic’s Claude Enterprise becomes more deeply embedded in workflows that go far beyond simple chat. According to Cisco’s announcement, the company is linking its AI Defense system to Anthropic’s inference hooks so that governed prompts can be checked before the model runs. The aim is to stop prompt injection and jailbreak attempts at the point where they matter most: the moment a user request reaches the model.
That matters because Claude Enterprise, Claude Code and Claude Cowork can act on a user’s behalf, reading documents, calling tools and executing tasks. Cisco says that makes the prompt itself an attack surface. The concern is not only direct malicious instructions from a user but also indirect attacks hidden inside shared files, tool outputs or other untrusted content that an agent may later consume. Anthropic’s own guidance on jailbreak defence and Claude Code security underlines the same risk, recommending input screening, careful handling of untrusted tool content and restrictions on risky commands.
Cisco’s pitch is that traditional data loss prevention is not enough for this environment. DLP tools are designed to spot sensitive data leaving an organisation, but not to understand when a model is being manipulated. AI Defense, by contrast, is built to inspect intent as well as content. Cisco says it can evaluate prompts, tool calls and conversation transcripts for prompt injection, jailbreaks and attempts to exploit tools, including those connected through the Model Context Protocol. The company says this lets it block malicious activity before inference starts, while still allowing safe requests to proceed.
The integration also fits into a broader AI security platform. Cisco describes AI Defense as covering discovery, inventory, validation, red-teaming, supply chain and model scanning, as well as runtime protection. That wider approach is meant to secure not just deployed prompts but the full AI lifecycle, from identifying agents and tools to enforcing policy at runtime. Cisco says the Claude Enterprise link is one enforcement point inside Cisco Cloud Control, its unified platform for the agentic era.
There are practical limits, however. Cisco says the current integration enforces checks on prompts because Anthropic’s inference hooks presently fire before model execution. Response-side inspection may follow if Anthropic extends the hook to cover outputs, and Cisco expects the enforcement path to become a native part of its Inspect API over time. For now, the company says the feature works with Claude Enterprise in beta, giving security teams a live inspection point without changing how employees use Claude. In Cisco’s view, that turns a binary trust decision into a governed control, allowing enterprises to keep the productivity benefits of agentic AI while adding policy enforcement against model-level attacks.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
