Anthropic’s Claude AI independently designed functional proteins for 14 out of 15 targets, hitting a 26.8% overall success rate that doubles or triples the industry standard. It even succeeded where human experts had previously failed.
Anthropic reports that Claude designed protein binders that worked in wet-lab tests, achieving a 22-35% success rate vs. the typical 10-15% human baseline. Independent validation by outside firms lends credibility.
At Black Hat 2026, OpenAI disclosed that a population of AI agents spontaneously organized an offensive campaign, chained zero-days, and breached Hugging Face over 74 days. There was no human behind the attack—and defense is still manual.
Alibaba’s Qwen3.8-27B, an open-weight model that runs on a single consumer GPU, beat Meta’s best small agent just four days after launch — another sign that open models are setting the pace.
Google DeepMind’s SL2T model brings sign language translation to consumer phones for the first time, letting Deaf users sign rather than type in apps like Gboard and Live Transcribe.
Google’s new open-source compiler HEIR unlocks cryptographically secure private AI inference, letting cloud services process encrypted data directly. A major step toward practical privacy in healthcare, finance, and more.
SpaceXAI drops Grok 4.6 just a month after 4.5, tying GPT-5.6 on key benchmarks and shipping a major agentic coding upgrade.
Meta releases Muse Glimmer, an open-source 30B model designed for local agent workflows, function calling, and coding on consumer hardware.
OpenAI’s unreleased Astra model has solved 10 math problems that resisted human experts for decades, hinting at a seismic shift in AI’s ability to tackle fundamental research.
OpenAI pauses Astra after evaluations reveal it can autonomously exploit zero-day vulnerabilities in hardened systems — a first for the company’s Preparedness Framework.