AI Briefing

Today's digest is about AI hitting the real world. Meta says its model co-wrote math papers that crack open problems, while Anthropic's Mythos turned a sloppy random number generator into an exploit that attackers weaponized in a day. Meanwhile Google is both launching chips into orbit and quietly locking its best Gemini models behind a paywall.

01ResearchMeta AI Research ↗

Meta's Muse Spark Co-Authored Six Math Papers, and Five Answer Open Questions

Meta published six papers written with human mathematicians and its Muse Spark model, five of which address previously open problems spanning probability, differential equations, group theory, optimization and arithmetic physics, while the sixth disproves a recent algebra conjecture. It matters because the work is transparent about which passages were AI-drafted and had a separate review team, a template for how AI can plausibly assist real research. Caveats remain: other teams independently solved some of the problems, and one paper leaves behavior at certain thresholds unresolved.

02IndustryThe Register ↗

Anthropic's Mythos Found a Math-Based Flaw in HFS, and Attackers Exploited It Within a Day

Anthropic's bug-hunting model Mythos spotted that Rejetto HFS (CVE-2026-61500) signs sessions using Math.random(), whose V8 algorithm is reversible, and leaked enough outputs to let a Z3 solver recover the seed and forge admin cookies for remote code execution. Horizon3 disclosed it on Wednesday and VulnCheck saw China-hosted exploitation by Thursday evening. The gap between AI-found bug and real-world attack is now measured in hours, and the fix is HFS 3.2.1.

03ProductAndroid Headlines ↗

Google Is Pulling Flash and Pro Models From Free Gemini Users on October 9

From October 9, free Gemini users are limited to Flash-Lite, AI Plus subscribers lose Pro, and Pro access is reserved for Pro and Ultra plans, while Deep Think expands from Ultra-only to Pro subscribers. Coming right after Gemini 4 Argon, it shows the top-end models are becoming the paywalled product as labs race GPT-6 Astra and Claude Opus 5.5. Expect the best AI to increasingly cost money.

04IndustryHardware Busters ↗

Google Put Four Trillium TPUs in Orbit to Test Whether AI Chips Can Survive Space

Project Suncatcher launched on October 1 aboard SpaceX's Transporter-18, carrying four Trillium TPUs and about 1 kW of solar power to see if they survive launch vibration, radiation during live workloads, and vacuum cooling. Ground tests with a proton beam simulating five years in orbit showed no hard failures, though HBM had uncorrectable errors Google deems acceptable for inference. A 2027 mission will test laser links between satellites, a step toward orbital data centers.

05PolicyBenzinga ↗

Appeals Court Pauses Minnesota's AI 'Nudification' Ban After xAI Challenge

The 8th Circuit temporarily blocked Minnesota's law, effective August 1, banning tools that create realistic intimate images of real people without consent, after a lower court had refused. xAI argues the law violates the First Amendment and says Grok Imagine has rigorous protections. The case will shape whether states can regulate generative image tools, and Attorney General Keith Ellison vows to keep defending it.