This week in AI research: Fields medalist says GPT-5.5 Pro did PhD-level math in an hour, Anthropic teaches Claude to ‘dream’

The past week or so, Anthropic’s gated Mythos model remains a news fixture, drawing White House policy attention, a Mozilla proof point and a somewhat skeptical review from an open-source project. Earlier this month, OpenAI had formalized a split between general-purpose and cyber-specific model access while Anthropic opted for a more reserved gated release for its Mythos model. In other news, Google flagged real adversaries using AI-augmented exploit research. On the math side, Fields medalist Timothy Gowers reported that GPT-5.5 Pro produced PhD-level additive number theory research in about an hour, while Google DeepMind published an agentic co-mathematician workbench. Alphabet announced a range of news: It launcheed Googlebook, and announced Gemma 4 inference optimization. Meanwhile, its Isomorphic Labs spinoff raised $2.1 billion for AI drug design. Finally, Microsoft supplied demand-side AI data.

Here are 10 items worth tracking.

1. Anthropic’s Mythos gets White House attention, a Mozilla proof point and some curl-project pushback: The White House is preparing an AI security order that would expand cybersecurity information-sharing programs to include AI companies. It would also address threats from more capable models. The news follows the March exit of David Sacks, who as White House AI & Crypto Czar, pushed hard for minimal federal AI regulation. Bloomberg reported that an earlier framing included a vetting system for models such as Anthropic’s Mythos, with NEC Director Kevin Hassett comparing the idea to drug approval. (Bloomberg)On the technical side, Mythos Preview lshined in Mozilla’s Firefox work, where Mozilla credited it with helping identify 271 vulnerabilities fixed in Firefox 150. But Daniel Stenberg’s curl recap told a different story. After scanning curl, one of the most heavily fuzzed and audited C codebases in open source, Mythos produced one confirmed low-severity vulnerability from five claimed findings.
2. OpenAI launches Daybreak for AI-assisted cyber defense: OpenAI’s Daybreak pushes frontier models into vulnerability discovery, patch generation and remediation verification. The system combines GPT-5.5, Codex Security and tiered access for verified defensive workflows. Daybreak offers secure code review, vulnerability triage, malware analysis and patch validation. The push also formalizes a split between general-purpose and specialized cyber access, with stronger account controls for high-risk security tasks. (OpenAI)
3. Google flags AI-augmented vulnerability exploitation: Google Threat Intelligence Group reported adversaries using AI-augmented research to identify and exploit a 2FA bypass vulnerability in a server administration tool. Google’s report also describes threat actors using middleware projects to aggregate API keys and support persistent access, which makes agent permissions, identity and tool access the main failure surface. (Google Cloud)
4. GPT-5.5 Pro produces ‘PhD-level math research’: Fields medalist Timothy Gowers reported that GPT-5.5 Pro solved multiple open problems from a recent additive number theory paper by Mel Nathanson. The model produced what Gowers called “PhD-level research in an hour or so, with no serious mathematical input from me.” Gowers noted that LLMs have reached the point where, if an open problem has an easy argument that human mathematicians missed, there is a good chance the model will find it. Open problems from recent papers used to be valuable training ground for early-career researchers, and that bar just got raised. Separately, a customized GPT-5.5 variant had helped discover a new Lean-verified proof related to asymptotic properties of off-diagonal Ramsey numbers. (Gowers’s Weblog)
5. Google DeepMind researchers publish an AI co-mathematician: Google and Google DeepMind researchers posted a paper introducing “AI co-mathematician,” an agentic workbench for mathematical research. The system is designed around the actual workflow of open-ended math: ideation, literature search, computational exploration, theorem proving, failed hypothesis tracking and theory building. The paper reports state-of-the-art results on hard problem-solving benchmarks, including 48% on FrontierMath Tier 4, and early use cases where the system helped researchers solve open problems or find overlooked literature. (arXiv)
6. Anthropic scales agent infrastructure with “dreaming” and a SpaceX compute deal: Anthropic introduced “dreaming,” a memory-refinement process for Claude Managed Agents that reviews prior sessions, finds patterns, updates memory and helps agents improve across jobs. Anthropic says Harvey saw roughly 6x higher completion rates in tests, while Netflix is using multiagent orchestration to process logs from hundreds of builds. Dreaming is still a research-preview feature, while outcomes and multiagent orchestration have moved into public beta. (Claude) On the compute side, Anthropic signed an agreement to use all capacity at SpaceX’s Colossus 1 data center. The collaboration adds more than 300 megawatts and more than 220,000 NVIDIA GPUs within the month. (Anthropic)
7. Isomorphic Labs raises $2.1 billion for AI drug design: Alphabet-spinoff Isomorphic Labs raised a $2.1 billion Series B to scale its AI drug design engine, expand globally and advance its drug candidate pipeline. The round links DeepMind’s research lineage, Demis Hassabis’ Nobel-era credibility, Alphabet/GV backing and pharma partnerships into a commercial AI-drug-discovery platform. (R&D World)
8. Google unveils Googlebook as a Gemini-first laptop category: Alphabet is turning Gemini into an operating-layer R&D program across hardware, Android, ChromeOS and personal computing. Google says Googlebook is built for Gemini Intelligence, with a “Magic Pointer” developed with the Google DeepMind team. It also offers custom Gemini widgets and tight Android integration. (blog.google)
9. Microsoft publishes its Q1 2026 Global AI Diffusion Report: Microsoft’s Global AI Diffusion Report found that AI usage rose from 16.3% to 17.8% of the world’s working-age population in Q1 2026, with 26 economies now above 30% adoption. AI adoption is accelerating, but unevenly across countries, sectors and labor markets. (The Official Microsoft Blog)
10. Google releases multi-token prediction drafters for Gemma 4: Google released multi-token prediction drafters for Gemma 4, saying the speculative decoding approach can deliver up to 3x faster inference without degrading output quality or reasoning logic. Inference efficiency is a growing constraint for local models, edge deployment, agentic workflows and cost-controlled enterprise systems. The release also extends the Alphabet thread running through the week: Gemini at the frontier, Gemma for open and local deployment, Googlebook for AI-native devices, Isomorphic for life sciences. (blog.google)