Google's AI Agents Prove Signal's Encryption Code Keeps Its Promises

Read the episode (Narration script)

Good morning and welcome to Cyber World News. The lead today is good news about security, for a change. Google DeepMind says its AI agents have mathematically proven that the code protecting your Signal messages does what it promises. After that, a less flattering scorecard for AI bug hunting, a computer scientist watches a machine settle his own conjecture, Claude starts attending meetings, and Toyota revives Scion to sell a truck that tops out around twenty five miles an hour. Let's get started.

Every message on Signal, and every chat in Google Messages, is protected by the same cryptography, the Signal protocol. Its design has been studied for years. Trouble usually shows up in the gap between the design on paper and the code that runs on a device. On Tuesday, Google DeepMind said a system called AlphaProof Code has closed a good part of that gap.

Human experts wrote down Signal's security goals as precise mathematical statements. If someone steals your keys today, they still can't read yesterday's messages. Once the break in ends, the conversation heals. Then teams of 40 to 80 agents running on Gemini Carbon wrote proofs that the real compiled program, the actual machine instructions, meets those goals. The proofs run past six million lines of Lean, a language where a computer checks every step, so nobody has to take the AI's word for it. The whole effort took about four weeks. The paper points out that verifying seL4, a famously secure operating system kernel, took twenty person years.

There's a nice sanity check too. The team put two old, long fixed Signal bugs back into the code, and both times the system proved the broken version was insecure. Signal's chief technology officer, Ehren Kret, called it "a major breakthrough."

The fine print matters. You still have to trust about 26,000 lines of human reviewed definitions, and the proof assumes the underlying ciphers hold up. It covers one to one conversations, with group chats left for later, and it covers the protocol's core library, not the whole app on your phone. Attacks that work by watching timing or power use are out of scope. Even so, DeepMind says this is the first time the running code of a widely used encryption protocol has been verified by any method. Finding a bug tells you about one bug. A proof tells you a whole category isn't there. DeepMind is publishing the proofs, and names medical device firmware and power grid software as places this could go next.

The other approach had a rougher day. Anthropic's Project Glasswing has spent seven months pointing its models at open source software and reporting what they find. Patrick Garrity of the security firm VulnCheck keeps score, and his latest tally is unkind. Anthropic's public ledger lists nearly 38,000 findings. 217 are marked fixed. Since early September the list grew by almost 12,000 findings and 15 fixes. The curl project received 45 reports. Of the 34 its maintainer has published so far, 19 were rejected as not vulnerabilities and 6 were confirmed, all low severity. Anthropic describes the reports from its newer self service scanner as "fully model generated, without human review or triage." Garrity's summary is that the bottleneck "didn't go away. It moved to the maintainers." VulnCheck sells vulnerability data, so it has a horse in this race, and plenty of those findings may be real and waiting on volunteers. Either way, machines now write bug reports faster than people can read them.

In 2009, Scott Aaronson and Andris Ambainis posed a conjecture about the limits of quantum speedups. A startup called Physical Superintelligence now has a proof, and TechCrunch reports that David Turturean, a recent MIT graduate there, got it using early access to Anthropic's Fable 5.2. It closes off one route by which quantum computers might have beaten ordinary ones, a route researchers already doubted. Aaronson says he hasn't yet understood the proof of his own conjecture. He believes it because a computer checked it in Lean and because researchers at MIT and Stanford wrote up an explanation. He lists his feelings as satisfied, awestruck, disappointed that "humans didn't get to snag this conjecture before the robots did," and fatigued "by the endless wonders from Aladdin's Cave." Turturean told TechCrunch he may never publish again.

Claude can now sit in on your meetings. Anthropic's Claude Meetings, in beta for business customers on Google Meet, joins a call as a visible participant, drafts the plan while people are still talking it through and answers questions in the chat. It won't join calls with outside guests unless an admin allows it, and if one shows up, it leaves. A participant that knows when to exit a meeting. That's new.

Federal regulators issued a construction permit for four small nuclear reactors at a Dow chemical plant in Seadrift, Texas, the first for a reactor on an American industrial site. They're X-energy's helium cooled design, 80 megawatts of electricity each, plus heat for the plant. The Nuclear Regulatory Commission hit its 18 month review target, which a critic at the Union of Concerned Scientists says it managed by pushing safety questions to a later date. Dow hasn't committed to building, and nothing would switch on before the early 2030s.

A few more. N-able, whose software lets IT shops remotely manage other companies' computers, took its hosted service offline after a researcher published dozens of unpatched flaws without warning. N-able says it has no evidence anyone has been broken into. A leak reported by 9to5Mac points to an unannounced Apple app called Harbor that would let you cable a few Macs together and pool their memory to run large AI models at home. And Toyota is bringing back Scion to sell a tiny Japanese style kei truck in America, starting with fleets in late 2027. It has four wheel drive, air conditioning, and a top speed of about 25 miles an hour, because that's the legal category it fits in.

That about wraps it up for today, friends. Thanks for listening; see you tomorrow!