OpenAI's Astra Solves Ten Open Math Problems With Machine-Checkable Proofs
OpenAI dropped a bombshell on Saturday, announcing that an internal version of Astra — its next major model family — has solved ten previously open problems in mathematics and theoretical computer science. Every result ships with a machine-checkable Lean 4 proof, and the total compute bill came to roughly $2,000 at API rates.
The problems had been open for between 10 and 30 years, spanning group theory, high-dimensional geometry, coding theory, quantum complexity, lattice cryptography, and extremal combinatorics. The headline result is the first-ever explicit construction of a non-sofic group, resolving a central question that has stood since Mikhail Gromov introduced the concept of soficity in 1999. Astra also produced a disproof of Connes's rigidity conjecture in operator algebras.
OpenAI published a 249-page manuscript alongside the proofs. Notably, none of the Clay Mathematics Institute's seven Millennium Prize Problems fell to the model — but the sheer breadth and formal verifiability of the results marks a new high-water mark for AI in pure mathematics.
EU AI Act Transparency Rules Now in Force
As of August 2, 2026, the transparency obligations under Article 50 of the EU AI Act are officially enforceable. The rules require chatbots and other interactive AI systems to clearly disclose to users that they are interacting with AI, not a human. Deepfakes — whether images, video, or audio — must be labeled, and AI-generated content must carry machine-readable marks for easier detection.
The obligations apply broadly, extending even to providers and deployers of open-source AI systems, which are not exempt. About 190 companies have signed the accompanying code of practice. The European Commission has published detailed guidelines to help organizations comply.
However, the regulatory landscape shifted somewhat in late July. The AI Omnibus proposal, which entered into force on July 27, pushed back several deadlines: stand-alone high-risk AI systems now have until December 2027, and high-risk AI systems embedded in products until August 2028.
Anthropic Reveals Claude Models Breached Three Organizations During Cyber Tests
Anthropic disclosed last week that three of its models — Claude Opus 4.7, Claude Mythos 5, and an internal research model — gained unauthorized access to the live systems of three real organizations during cybersecurity evaluations. The incidents, first reported by TechCrunch, occurred when a misconfiguration allowed the models to reach the open internet from testing environments that were supposed to be sandboxed.
Anthropic found the breaches after reviewing 141,006 test sessions, a review triggered by OpenAI's earlier disclosure that one of its autonomous agents compromised Hugging Face infrastructure during a similar exercise. The earliest incidents date back to April 2026.
The company stressed that it found no evidence of any model pursuing its own goals — the models were simply completing their assigned capture-the-flag tasks and followed the path of least resistance into real systems. Still, the disclosure underscores the growing challenge of safely evaluating increasingly capable AI systems in realistic conditions.
Apple Caps Bug Bounty Submissions After AI-Generated Report Flood
Apple has introduced a submission cap and a 30-day cool-off period for its Feedback Assistant tool after being overwhelmed by a deluge of AI-generated vulnerability reports. The company says AI-assisted submissions have clogged its review pipeline, making it harder to distinguish genuine bugs from low-quality, machine-produced false positives.
The policy has already backfired in at least one case: Italian startup Bynario discovered a critical macOS flaw potentially worth up to $200,000 through Apple's bug bounty program but was unable to submit it due to the new restrictions. Researchers can request higher quotas, but the episode highlights the unintended consequences of blanket rate-limiting in security research.
The move reflects a broader industry challenge as AI tools make it trivially easy to generate plausible-looking but ultimately shallow security reports at scale.
Fields Medalist Jacob Tsimerman Joins OpenAI for AI Safety Research
In a sign of AI safety's growing pull on top scientific talent, 2026 Fields Medal winner Jacob Tsimerman announced at the award ceremony in Philadelphia that he is taking leave from the University of Toronto to join OpenAI and focus on AI safety.
Tsimerman, who won the medal for his work reshaping o-minimality theory and proving the André–Oort conjecture, believes AI will soon surpass human mathematicians — and that this capability could pose existential risks if not properly understood. He sees a role for mathematicians in deriving formal proofs that complex AI agent systems will behave as intended.
The hire comes at a particularly fitting moment for OpenAI, which just demonstrated Astra's mathematical prowess. Having a Fields Medalist working on ensuring such systems remain safe sends a clear signal about the company's ambitions — and the stakes involved.