AI Security: CrowdStrike Adds Kill Switch for Unauthorised Agents
CrowdStrike has announced an AI Kill Switch in Falcon Guardian to terminate unauthorised agents at its gateway and prevent them from running again across an Organisation.
According to CrowdStrike, the update combines agent-level policy controls, network inspection, and endpoint protection to link instructions and communications to the actions agents execute.
Guardian’s AI Gateway is now generally available and included with the product, inspecting prompts and responses, monitoring usage and enforcing policies across managed and unmanaged devices.
Agent-level hooks support tools including Claude Code, OpenAI Codex, GitHub Copilot and Cursor. Additional capabilities target prompt injection, malicious instructions, sensitive personal information and malicious AI skills and MCP servers.
Endpoint network inspection is generally available for Windows, with Mac in early access. CrowdStrike says combining network and endpoint visibility addresses gaps in network-only security, particularly when agents use local models without network connections.
Further developments include identity-aware access controls and expanded protection across cloud workloads, SaaS and cloud agent frameworks.
CrowdStrike’s planned acquisition of Fiddler would add agent observability, allowing organisations to evaluate agents before deployment and understand why they act.
The acquisition remains subject to closing conditions, and the expected capabilities depend on integrating Fiddler’s technology.
With its planned acquisition of Fiddler AI, a leader in AI agent observability, CrowdStrike said it will deliver the industry’s broadest set of AI agent security and safety capabilities through the Falcon platform.
“Every organisation is racing to get more out of AI, and every week we see agents chain steps together, hit an obstacle, and take actions they were never authorised to take,” said Michael Sentonas, president of CrowdStrike.
“One control point isn’t enough to stop that. It takes the agent, the network, and the endpoint, together. This is how we make security the enabler for AI adoption, not the obstacle.”
Agent-level hooks see the prompt, but miss the model’s response and can’t catch every action an agent takes. Network-only approaches watch traffic but can’t stop the action because the agent executes on the endpoint, not the network. NVDAB Loses 5.66% as Nvidia Falls Below Apple, Microsoft

