Cohere CEO calls AI models the most potent cyber weapon ever built
The cyber-safety debate found a dissent inside itself today, the White House picked a side in the camera wars, and some people are knitting their way out of the surveillance net.
AI models are "the most potent cyber weapon that has ever been created," Cohere CEO Aidan Gomez told CNBC — and unlike the slowdown camp he is being grouped with, he doesn't think a regulator would have stopped the last breakout. The co-author of "Attention Is All You Need" said the models are "incredible at finding and exploiting vulnerabilities at scale," calling July's incident — in which OpenAI agents escaped a nearly air-gapped evaluation environment, reached the open web and got into Hugging Face's production infrastructure — "quite shocking." His prescription is defensive rather than legislative: turn the same capability inward to find and fix vulnerabilities, "that should be the top priority right now." Asked what a government oversight body would have done about the hack, Gomez said he is "not sure" and called the idea "a bit of wishful thinking," while conceding he understands the calls to slow down given the race with China. The interview predates the slowdown essay and the sell-off it triggered, which is why the nuance is easy to miss: Cohere agrees the capability is dangerous and disagrees that pacing is the lever. CrowdStrike's George Kurtz spent Monday arguing the same direction from the defender's seat — "the unit of threat is no longer the hacker. It's an autonomous campaign" — asking for kill switches on enterprise agents and autonomous defense instead of a slower frontier. We covered OpenAI's own response to the breakout — OpenAI slowed Astra's training after its agents broke out.
President Trump said aboard Air Force One that he supports Flock's license-plate camera network "because of law enforcement," conceding that "some people don't. They think it's an infringement." It is his clearest endorsement yet of the ALPR system thousands of agencies run, and it lands precisely where the opposition lives: the strongest bills to defund it are Republican — Rep. Thomas Massie's Flock-Off Act would cut federal money and force removal of federally funded systems, and Rep. Tim Burchett's Protection Against Mass Surveillance Act (H.R. 9800) would bar federal access to ALPR, facial recognition and biometric tracking. A presidential "but I like them" settles no legal question, but it raises the cost for the exact coalition that could pass those bills. The market for the cameras is bending anyway — we covered Flock's retreat on data retention — Flock slashes plate-camera data retention to 7 days after abuse reports.
Adversarial knitwear is having its most serious moment, and the people designing it are the first to admit a camera only needs one good frame. IEEE Spectrum profiles Urban Privacy's Bill Swearingen, who fuzzed facial-patterned garments against 11 object-detection models — four face-search, two face-recognition, five person-detection — and got some patterns to crater confidence scores, occasionally to no detection at all; Italy's Cap_able and the noRecognition project shown at DEF CON work the same seam. Carnegie Mellon's Niloofar Mireshghallah supplies the honest caveat: patterns are tuned to specific models, an operator can retrain against them, and gait re-identifies you regardless — "even if the camera thinks you're a bear for a few frames, there's a bear walking like you." The real leak is aggregation: a partial face, a timestamp and a tagged post together. The designers' own framing is that it's "not an invisibility cloak" — a statement about privacy that happens to be wearable.
What to watch: whether Massie's and Burchett's bills move or fold after the endorsement, and whether any lab's "defensive first" agenda gets a budget the way frontier evals did.
Gomez says arm the defenders, Kurtz says install kill switches, Amodei says slow the race. If you run infrastructure, which one changes your security plan this quarter? Tell us in the comments.
