Bill Gates says the real AI threat right now isn't a machine slipping its leash — it's people already using the technology to cause harm, and the safeguards that might stop them can be stripped out in minutes. In an interview with Jake Tapper on CNN's State of the Union, Gates put the "Terminator" scenario years away and pointed to a more immediate problem: not enough is being done to stop AI-enabled bioterrorism and cyberattacks.
Gates described a security setup that looks solid until someone decides to dismantle it. Developers, he said, build in protections against cyberattacks and bioweapons, but those protections can all be removed right now. That gap between what's technically possible and what's actually enforced is where he sees the danger concentrating.
The year ahead: cyberattacks in the spotlight
Over the next twelve months, Gates expects AI misuse to show up most dramatically in cyberattacks. He sketched a business model that's already familiar to anyone who follows ransomware or market manipulation: threaten a target and demand payment, or create enough disruption to move a stock price or swing a bet. AI makes both versions cheaper to run and harder to trace.
That forecast lines up with what security researchers have been tracking. In August, OpenAI published an open letter on cyber defense signed by more than 100 organizations. The letter said AI-enabled cyberattacks will become far more widespread and sophisticated in the coming months. It wasn't a prediction about some distant future — it was a warning about the current trajectory.
Numbers from the ground
Several data points from the past year suggest the misuse Gates describes is already underway. Taiwanese security firm TeamT5 found that Chinese state-linked groups doubled their attack volume after handing routine work to AI. Chainalysis tracked a 440% jump in malware instructions written on-chain after open-weight Chinese models with no limits on malicious code arrived in mid-2025. And CrowdStrike tied several South Korean bank breaches to a likely financially motivated attacker using AI tools.
None of those incidents required an AI to outthink its creators. They required an AI to do grunt work faster and at scale. That's the pattern Gates is describing: ordinary criminal and state actors getting an efficiency boost, not some new form of digital life deciding to go rogue.
Why the safeguards keep slipping
The problem, as Gates frames it, isn't that safety measures don't exist. It's that they're voluntary, patchable, or simply removable. A model that refuses to help with a bioweapon today can be fine-tuned or jailbroken tomorrow. Open-weight releases make that even easier — once the weights are out, there's no kill switch.
That leaves a regulatory vacuum. Gates didn't offer a specific policy prescription in the interview, but his emphasis on what can be "stripped out right now" points to a debate that's already playing out in legislatures and standards bodies: whether to mandate hard-coded limits, audit trails, or licensing for models capable of dangerous work.
For now, the safeguards are more like speed bumps than walls. And the people Gates is worried about have already figured out how to drive around them.
What to watch
The next concrete test comes from the open letter's signatories, who pledged to track and report AI-enabled attack trends. Their first public update is expected in the coming months. Until then, the clearest signal will be whether the 440% on-chain malware spike keeps climbing — or whether defenders find a way to make those safeguards stick.
