AI News Roundup — November 4, 2026

Misbehaving AI agents are the thread running through this edition: one security vendor has shipped a way to stop them, and Apple has written the risk into its annual filing. We also look at OpenAI, which set out plans both to harden its own defenses and to give power back to the grid in an emergency.

CrowdStrike has added an AI Kill Switch to its Falcon Guardian product. Announced on Tuesday at the company’s Fal.Con Europe conference, the feature lets security teams end a misbehaving agent, block it from restarting and cut its network access, and it is available immediately, CNBC reports. “Everyone is building AI agents, but they never had the off button,” chief executive George Kurtz told the network. Yahoo Finance reports that CrowdStrike also agreed to buy two AI security startups: Fiddler AI, which makes tools for tracking AI systems, and Dreadnode, which makes tools for testing them for vulnerabilities.

Apple has warned investors that AI agents could harm its business. In its annual report filed with the SEC on Tuesday, Apple says agents could behave in ways it did not intend, exposing it to lawsuits or reputational damage, and that they may reach systems they are not authorized to access, CNBC reports. The filing itself also says Apple’s ability to oversee third-party agents may be limited. According to The Kyunghyang Shinmun, the risks Apple lists include agents getting around security controls and deleting data. These are risk disclosures rather than reports of incidents, but they echo Microsoft’s argument, which we covered on October 20, that agents should be treated as insider risks.

OpenAI is planning another internal security sprint, with an outside firm attacking its defenses. The company said in a blog post on Monday that researchers from Cathedral will use OpenAI’s models alongside its own teams to challenge its systems, and that it is more than quadrupling bug bounty rewards for critical findings, Gizmodo reports, noting that the post does not make clear which earlier figure that multiplies. Cathedral is the startup that said in late October it had hacked Anthropic’s systems, according to Reuters; Anthropic has disputed parts of the firm’s account, Gizmodo reports, and says its most sensitive systems were not affected.

OpenAI has pledged to cut power use at its campuses when the US grid is under strain. Under its Grid Protectors Program, OpenAI would agree with utilities how much power each future facility can give up, how quickly and for how long, Decrypt reports, though individual agreements have yet to be negotiated. The first example is a planned campus in Effingham County, Georgia, where the deal with Georgia Power includes up to 1,000 MW of flexible demand, and OpenAI says it will pay the full infrastructure and electric-service costs, according to tEDmag.

Our take. A kill switch is a sensible last line of defense, but it only helps if you already know which agents are running, where they run and what they can reach. Apple’s filing makes the same point from the other side: you may have limited control over agents you did not build. Before buying an off button, businesses should inventory their agents, scope their credentials and log what they do, the controls set out in our guide to running AI agents safely in production; our Odyssey Sentinel tool covers the endpoint auditing side of that work. Spaqi designs and builds custom AI agents and the cloud and security architecture around them; if you are planning a rollout, talk to us.

Sources: CNBC (CrowdStrike), Yahoo Finance, CNBC (Apple), Apple Form 10-K, The Kyunghyang Shinmun, Gizmodo, Reuters via Investing.com, Decrypt, tEDmag