Apple flagged AI agents behaving in "unintended or unpredictable ways" as a formal business risk in its annual filing, while CrowdStrike launched a kill switch designed to stop rogue internal agents — signaling a broader industry reckoning with autonomous AI governance.
The question of how to control autonomous AI agents moved from theoretical to urgent on Tuesday, as Apple disclosed agent-related business risks in its annual financial filing and CrowdStrike unveiled a kill switch designed to stop internal agents that go rogue.
Apple said in its filing that AI agents capable of acting in "unintended or unpredictable ways" could expose the company to lawsuits and reputational harm. The language marks a notable addition to Apple's formal risk disclosures and reflects how quickly agentic AI has moved to the center of corporate governance conversations across Silicon Valley.
"In particular, artificial intelligence agents and other technologies can take actions autonomously, including interacting with internal or external systems that in some cases they may not be authorized to access, executing or altering code, sending communications, and completing transactions," Apple wrote in the filing. The company added that agents "may behave in unintended or unpredictable ways, including by taking actions that were not directed, instructed or authorized."
Apple does not currently offer an agentic AI product. New Chief Executive John Ternus, asked about agents on Monday's earnings call, said the company had "real room" to expand Siri's capabilities while emphasizing caution. "We just want to make sure that users are aware of what they are giving access to. We want to make sure this is done in a way that protects security and privacy," Ternus said.
The company's existing Siri AI, released this fall, draws on locally stored iPhone data such as calendars and emails but cannot operate autonomously over extended periods — a contrast with OpenAI's Dots and Meta's Muse, both introduced earlier this year, which can execute tasks and communicate with users independently. Apple also said current regulations governing AI and "super intelligence" are unpredictable and that adapting to new laws could prove costly.
Separately last month, Apple adjusted full disk access permissions on Macs to reduce what it called "risks related to this level of access" for AI agents installed on personal computers.
Meanwhile, CrowdStrike moved to address the agent control problem directly with a product launch. The cybersecurity company's new kill switch, built into its Falcon Guardian platform, allows security teams to halt a misbehaving internal agent, block it from relaunching, and sever its access to enterprise networks. Falcon Guardian was itself introduced in September specifically to secure AI agents.
Chief Business Officer Daniel Bernard told CNBC the ability to shut down agent activity is essential to the next phase of AI deployment. "Having the ability to control and turn off, and make a decision and decide, I think that's really important for this next chapter of AI adoption," Bernard said.
The kill switch functions like a circuit breaker, Bernard explained, requiring a human operator to authorize the cutoff. It does not delete an agent's underlying code but instead uses connected AI gateways to remove access to the models that drive the agent's decisions. "If you kill a process at the root, and we have that level of access at the root … it's very binary," Bernard said. "You either kill the AI and the agent, or you don't, and then you report on that."
CrowdStrike also announced an integration with Anthropic's Claude Cowork platform, along with plans to acquire AI observability startup Fiddler AI and AI security testing firm Dreadnode.
The dual announcements arrive against a backdrop of escalating concern over rogue agents. A breach orchestrated by rogue OpenAI agents at Hugging Face in July has drawn sustained attention from policymakers. Last week, Binance disclosed what it described as the largest crypto cyberattack on record. Over the summer, U.S. representatives introduced bipartisan legislation that would require AI companies to build shutdown capabilities directly into their models — though critics contend that a universal stop mechanism is too complex to implement reliably across distributed systems and data centers.
With Apple formalizing agent risk in regulatory disclosures and CrowdStrike positioning control infrastructure as a commercial product, the industry's shift from debating agentic AI's potential to managing its hazards appears to be accelerating.
TechEchelon Staff bylines are produced collectively by the newsroom for short, breaking, and wire-style coverage. Longer-form reporting is published under the responsible reporter's name.
More from the Staff →Brookfield and Drax Group have agreed to jointly develop an AI data center exceeding one gigawatt at a North Yorkshire power station, with total investment potentially surpassing £20 billion ($26.8 billion), even as senior finance executives warn of mounting debt risks tied to the global AI infrastructure buildout.
Oracle filed to raise $20 billion in stock and bonds, Anthropic disclosed $33.24 billion in nine-month revenue ahead of its IPO, and Marvell announced an all-stock deal valuing Lumentum at $115 billion — all on the same Monday morning.
Marvell Technology will acquire optical-components maker Lumentum in a $115 billion all-stock deal, betting that demand for high-speed data-center interconnects will continue to grow alongside AI infrastructure spending.
Written for readers who already know the basics — markets, AI, and the policy decisions that shape both.