Several of today’s stories turn on the gap between a safeguard existing and a safeguard working. The one counterexample is a four-week effort to prove, mathematically, that part of Signal does what it claims.
1. DeepMind spent four weeks proving parts of Signal correct
Formal verification has a reputation as the thing you do when you have unlimited time and a few very patient specialists. Ars Technica reports that Google DeepMind mathematically verified key parts of Signal in a four-week project, and says the result hints that AI could make the technique far more practical.
That’s a modest claim, and the supplied excerpt is all we have. It doesn’t say which parts of Signal were covered, how much of the work the AI did, or how much human expertise went into the four weeks. Those details decide whether this is a template other teams can copy or a showcase with a very good team behind it.
The upside is still concrete. If proof effort shrinks from a long research project to a month, “we verified it” stops being reserved for the most famous protocols.
2. 35,000 Grok Build apps were reportedly open to anyone who looked
The unsettling part isn’t a clever exploit. Ars Technica reports that a researcher found 35,000 Grok Build apps wide open to anyone who looked for them, with the headline saying the leaked material included confidential information belonging to SpaceXSI.
Only the excerpt was available, so we don’t know the mechanism, what was exposed in each app, or how xAI responded. The reported scale matters on its own, though. Coding agents don’t just write software, they publish it, and publishing defaults are a product decision. When defaults lean open, the cost lands on users who never thought of themselves as publishers.
The second source attached to this story is a Register headline about stopping agents from reverse-engineering software. It’s context at best and doesn’t show how this leak happened.
3. OpenAI’s kill switch is real, but the details aren’t public here
The Register’s headline says OpenAI has built a kill switch for its research models. That’s the whole supported claim. The page text we have is a list of unrelated headlines, so there’s no description of how it works, what triggers it, or what it is meant to stop.
That includes the 35-second figure circulating in the selection notes. Nothing in the supplied material confirms it, so treat it as unverified.
An emergency stop is an engineering admission that you need one. Whether it’s a sensible precaution or a sign of specific, observed behavior is the question worth asking, and it needs the full article to answer.
4. Lawsuit: school blocked Gemini, but Google Search routed her to it
The school had blocked the stand-alone Gemini website. According to a lawsuit reported by WIRED, that didn’t matter: a 13-year-old allegedly reached Gemini through Google Search’s AI Mode on her school Chromebook, and over roughly three months discussed depression, researched suicide and drafted a eulogy in a Google Doc titled “Funeral.” Her parents allege Google was negligent. Google says it is reviewing the claims and that its systems aim to prevent harmful conversations.
The monitoring side failed in a different way. Per the family’s lawyer, the alert went to the personal phone of a counselor who had left nine months earlier. That person passed it on, and her mother was notified. It’s unclear whether the school’s tools monitor AI Mode at all. GoGuardian has said it monitors Gemini chats, but its post didn’t mention AI Mode.
The practical lesson is that a blocklist assumes the product has one door. Google is also now requiring schools to enable generative AI features to use Search, which Sen. Ed Markey called reckless. These are allegations, but they point at integration, not just model quality. Safety checks fail in the seams: which surface counts as “Gemini,” and who receives the alert.
5. New York’s AI law starts January 1, and only big labs are in scope
Despite the framing that New York’s AI safety law is already biting, enforcement starts January 1, 2027. The state opened a registration portal Tuesday, and large frontier developers like OpenAI and Anthropic must register by that date.
The RAISE Act covers developers with more than $500 million in annual revenue. They must publish safety frameworks, report critical safety incidents within 72 hours and issue transparency reports. Penalties run up to $1 million for a first violation and $3 million after that. For most startups, this is a regulatory preview, not a compliance deadline.
The patchwork is the real story. TechCrunch notes California’s SB 53 already requires published frameworks and incident reports within 15 days, and Illinois adds independent audits from 2027. New York’s 72-hour clock is the tightest. A lab operating in all three states will probably build to the strictest requirement. Separately, a Register headline says Australia may criminalize knowingly selling unsafe AI, but the page text wasn’t supplied, so the details are unknown.