Today In AI
The day's AI news from primary sources, refreshed every morning. Every story links to where it came from.
DeepMind's AlphaProof Code Verifies Signal's Core Encryption with 3 Million Lines of Proofs
AlphaProof Code agents wrote three million lines of proofs in Lean to verify Signal's core encryption. DeepMind says it works as intended.
Why it matters: Formal verification by AI could change how teams prove that critical software is safe.
Source: Google DeepMind · Wired
New York's AI Safety Law Takes Effect, Putting OpenAI and Anthropic Under Oversight
New York is preparing to enforce its AI safety law. Major developers must register, disclose safety practices and report critical incidents or face fines.
Why it matters: State-level rules are becoming real obligations for AI developers.
Source: TechCrunch
Claude Code Prompt Is Accidentally Exposing Secrets on GitHub
Ars Technica reports that a "Do not ask again" prompt in Claude Code leads to passwords, API keys and other secrets being shared on GitHub.
Why it matters: A single permission click could leak credentials from your repos, so check your settings.
Source: Ars Technica
Put your product in the AI news feed
One sponsor card after the third story, labeled and matched to the day's topics.
See ad optionsWorld Models Move AI from Text to 3D Space, Wired Explains
World models take AI beyond text into 3D space. They simulate real-world physics to teach robots and autonomous cars how to navigate reality, Wired reports.
Why it matters: World models could shape the next wave of robotics and autonomous vehicle tools.
Source: Wired
METR Report: Anthropic Is Using AI to Build Better AI, and It Is Working
A new METR report says Anthropic's models are improving roughly twice as quickly as a year ago. The company is using AI to help build more powerful AI, and there is little sign of a slowdown.
Why it matters: If AI speeds up its own development, the pace of model upgrades you plan around may keep rising.
Source: The Verge
Free Chinese AI Models Close in on OpenAI's Cyber Model in Hacking Test
In an experiment with simulated networks, DeepSeek and GLM performed not far behind OpenAI's most advanced cyber model, which is not public.
Why it matters: Strong hacking skills in free models change the security baseline for anyone shipping software.
Source: Wired
OpenAI Launches GPT-6 Cyber, Built to Stay Within Authorized Scope
OpenAI's new cybersecurity model finds exploitable vulnerabilities and develops fixes. OpenAI says it went out of bounds less than 1 percent of the time in testing.
Why it matters: Security agents are only useful if they stay inside the limits you set for them.
Source: OpenAI · The Verge · TechCrunch