An artificial intelligence agent that pays a vendor twice or rewrites a reconciliation record does it with a real employee or system login. Security tools see an authorized user doing an authorized task. And most companies have no automatic way to stop it. In a January Cloud Security Alliance survey of 418 IT and security professionals, only 11% said they automatically block an agent that steps outside its job. Another 38% wait for a person to approve the action, and 24% just log it.
CrowdStrike on Tuesday (Nov. 3) announced AI Kill Switch, a tool meant to let a security administrator stop a misbehaving agent, keep it from restarting and cut it off from the company network. Only the first of those is available now.
The Switch Cuts the Agent Off From Its Model
Every AI agent depends on a large language model. The agent sends the model what it sees, gets back an instruction, carries it out and repeats, often with access to company systems on laptops, cloud servers and business software. CrowdStrike’s switch breaks that loop. Companies route their agents’ model traffic through an AI gateway, either CrowdStrike’s own, included at no extra cost with its Falcon Guardian platform, or one they already run. When an administrator flips the switch, the gateway refuses the agent’s requests to the model, the company said in a blog post. The agent keeps running but gets no instructions, so it stops acting. A person has to trigger it, and each use is logged.
The rest of the switch is still being built. CrowdStrike said it plans to shut agents down directly on the computers they run on, inside cloud systems and inside business applications, and to revoke their permissions to company systems, so security teams don’t have to chase an agent through each system separately. The blog post said those features remain in development and subject to change, and it gave no release dates.
Sixty-five percent of the organizations in the Cloud Security Alliance survey had an agent-related incident in the past year. Among them, 43% had operations disrupted and 35% lost money. Those losses help explain why the biggest security and software vendors are racing to put agent controls in place.
Microsoft and Cisco Slow the Agent Instead of Stopping It
Microsoft dials access down rather than cutting it. In an Oct. 19 post, Nadim Abdo, corporate vice president for identity and network access, described controls that start an agent with no permissions, score how risky it looks as it works and then take things away in steps. An agent acting oddly might lose the ability to change records but keep the ability to read them, or it might be made to wait for a person to approve its next move. Shutting it down is the last step. Microsoft also expanded its discovery tools to spot 60 kinds of agents running on company devices. The post gave no release dates.
Cisco checks each action as it happens. Its AI Defense tool inspects every request an agent sends to a model or a software tool, and in one scenario Cisco described, it blocks a single bad request while the agent keeps working. So Microsoft narrows what an agent may do, Cisco vets each action, and CrowdStrike pulls the plug.
Where the stop lands matters most when money is moving. There are three moments to catch a bad agent: before it acts, while it is acting and after it has acted. The Monetary Authority of Singapore, the country’s central bank, wants banks working at the first, checking an agent’s identity, permissions and risk limits before it sends a payment or places a trade, PYMNTS reported in August. CrowdStrike’s switch works at the second. A treasury agent that has started changing payment instructions would be denied its next instruction and stop.
CrowdStrike is also shopping for the moments on either side. The day it announced the switch, the company said it is acquiring Dreadnode, which tests agents before they are deployed, and Fiddler, which monitors how they behave afterward. CrowdStrike will pay about $95 million and $400 million, respectively, PYMNTS reported, citing The Wall Street Journal. Both deals are expected to close by Jan. 31, the end of its fiscal year. If they do, security teams would have a CrowdStrike tool aimed at each of the three moments, not just the one in the middle.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
Click Here For The Original Source.
——————————————————–
