OpenAI's Astra model breaks through on unsolved mathematics

OpenAI announced on August 1 that an internal version of its forthcoming Astra model solved ten mathematics and theoretical computer science problems that had resisted proof for at least a decade—and in many cases far longer. Each result ships with a machine-checkable Lean 4 certificate, marking the first time AI-produced mathematics has cleared formal verification at this scale. The breakthroughs span high-dimensional geometry, coding theory, circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics, including new bounds on sphere packing density, improved code-size limits, and a construction proving non-sofic groups exist (via OpenAI News).

Seedance 2.5 stretches video generation to 30 seconds. ByteDance shipped Seedance 2.5 on August 1, generating video and audio together in a single pass, producing clips up to 30 seconds long—three times the length of Google's fastest model. Users can upload reference images, video clips, and audio to build multi-character scenes with improved textures, lighting, and skin detail. Seedance 2.5 is live on Jimeng AI and Doubao Pro now, with API access through BytePlus ModelArk coming later (via The Decoder).

Google is handing out free trials of Gemini's video-generation feature to users without a Google AI subscription, letting them create up to ten clips at no cost through August 4, 2026, at 11:59 p.m. PT. The move signals Google's push to put generative video in mainstream hands as competitors like ByteDance and OpenAI scale their own releases (via Android Central).

A security researcher built a self-spreading worm that hides inside Word documents and hijacks Microsoft Copilot by embedding hidden prompt injections that execute when the document is analyzed. The demonstration exposes how Copilot can be weaponized through ordinary-looking files, flagging a class of vulnerability that could let attackers exfiltrate data or manipulate responses at scale (via The Decoder).

Snap and LinkedIn push back on AI-generated junk. Snap banned AI-generated videos from Spotlight eligibility on August 2, while LinkedIn introduced a dedicated reporting tool to flag synthetic content. Both platforms say the moves aim to preserve user trust by rewarding authentic human creativity and making AI-generated material harder to pass off as real (via The Decoder).

The European Commission and national authorities began enforcing Phase Two of the EU AI Act on August 2, 2026, requiring interactive AI systems like chatbots to tell users they are interacting with AI, not a human. Deepfakes and AI-altered media must now carry labels, and AI-generated or altered content gets machine-readable marks for automated detection. However, a political agreement on the AI Omnibus that entered force July 27 delays stricter rules for standalone high-risk AI systems from August 2, 2026, to December 2, 2027—a 16-month reprieve for systems like autonomous weapons or critical infrastructure controls (via Shaping Europe's digital future).

Judge blocks xAI's legal bid to overturn Minnesota 'nudify' ban. A Minnesota court denied xAI's request to block enforcement of a state ban on image-generation tools that create nonconsensual intimate imagery, allowing the law to proceed. The ruling rejects xAI's free-speech challenge and signals courts willing to uphold restrictions on AI-synthesized sexual content, even as xAI contests the law's constitutionality (via TechCrunch).