Federal Agencies Test Claude Mythos for Cyber Defense as GPT-5.4 Pro Solves an Erdős Problem

Federal agencies are testing Claude Mythos for cyber defense, working around the White House's ban on Anthropic. Treasury is also pushing for access to use it for vulnerability hunting. The interest tracks with the UK AI Security Institute's findings: Mythos Preview solved 73% of expert-level capture-the-flag challenges (security puzzles where you find and exploit vulnerabilities in a controlled system) and was the first model to fully solve "The Last Ones," a 32-step simulated corporate network attack estimated at 20 hours of human work. It completed the full attack in 3 of 10 attempts and averaged 22 of 32 steps, compared to 16 for the runner-up, Opus 4.6.

Amazon Bedrock has added Mythos to a gated research preview. OpenAI released GPT-5.4-Cyber, a defensive variant built for the same security testing work.

Claude Opus 4.7 may release as soon as this week, separate from the Mythos line. Some users have reported that Opus 4.6 and Claude Code feel slower and less consistent recently, which tends to happen when compute gets shifted toward training the next model.

On the research side, Anthropic showed that a weaker model can fine-tune a stronger one (a setup meant to simulate humans overseeing AI smarter than them) and close 97% of the capability gap in days for about $18,000. This is a meaningful result for alignment, the field focused on keeping AI systems behaving the way we want as they get more capable.

GPT-5.4 Pro solved an open problem from mathematician Paul Erdős, specifically Problem #1196. Jared Duker Lichtman, a mathematician who reviewed the proof, called it elegant enough to belong in what Erdős called "The Book," his term for the most beautiful possible proofs of any given theorem.

Cosmo Pharmaceuticals reported one-year results for clascoterone, a topical drug for male pattern baldness. The drug kept hair growth going through the year-long extension study, which matters because most existing baldness treatments work on the whole body and come with hormonal side effects. A topical option avoids that.

Separately, Amazon launched Bio Discovery, a tool that lets researchers run drug discovery workflows with AI agents coordinating lab experiments, making this kind of work accessible without building the infrastructure in-house.

Uber has spent its entire 2026 AI budget a few months in, almost entirely on Claude Code, according to CTO Praveen Neppalli Naga. Anthropic also launched Claude Code routines, which are agents that run on a schedule or in response to specific events (like a code commit or a failed test) on managed cloud infrastructure.

Apple is moving 200 engineers from its Siri team to an internal AI coding bootcamp, a notable reallocation given where voice assistant work typically sits on the priority list.

Allbirds, the shoe company, sold for $39M and is rebranding as "NewBird AI," a GPU rental business. The stock jumped 582% on the announcement. Maine became the first state to ban construction of data centers over 20 megawatts, with the ban running through late 2027.

OpenAI reports that ChatGPT's user base, once about 80% male by first name, is now balanced. This is a common pattern when a technology shifts from early adopter to general purpose.

Elon Musk is pushing suppliers to accelerate his Terafab chipmaking project, sourcing equipment from Applied Materials, Tokyo Electron, and Lam Research. USC researchers built a memristor (a type of memory chip that also does computation) that works reliably at 700°C, hotter than the surface of Venus.

Meta committed to 1 gigawatt of custom AI chips with Broadcom built on a 2-nanometer process, which is the most advanced chip manufacturing node currently in use. Amazon agreed to buy Globalstar for $11.57B to expand its low Earth orbit satellite business.