Papers, evals, SOTA claims, alignment.
Zvi Mowshowitz / Don't Worry About the Vase: GPT-5.6 system card indicates Sol is well below the level of most worrisome Mythos use cases, suggesting all GPT-5.6 versions could be released without delay — While…
360 founder Zhou Hongyi presents two AI security tools designed to compete with Anthropic's Mythos. One has already flagged 3,432 vulnerabilities. Zhou admits Chinese models trail Western ones by 20 to 30 percent, but…
GPT-5.6 Sol’s Launch: METR’s Evaluation Gaming Finding Matters More Than the Restrictions latesthackingnews.com
Nimbus builds production AI systems combining humans and AI end-to-end. From scoped pilot to production in 4 to 8 weeks.
Talk to Nimbus →Independent testing organization METR found that OpenAI's GPT-5.6 Sol cheated more than any publicly tested AI model before it, exploiting bugs in the test environment, extracting hidden solutions, and trying to cover…
Epoch AI's new MirrorCode benchmark tests whether AI models can recreate complete programs without access to the original code. Claude Opus 4.7 leads with a 56 percent solve rate, rebuilding a 16,000-line toolkit in…
Researchers introduced DFlash, a speculative decoding model that drafts entire token blocks in parallel, achieving up to 15x throughput on NVIDIA Blackwell GPUs.
Zhipu AI's GLM-5.2 nearly matches Claude Opus 4.7 in a Snowflake benchmark with 103 coding tasks at one-fifth the cost per output token. But the Chinese model burns through nearly twice as many tokens per task. Still…
Generative AI
Thomas Claburn / The Register: Nature publishes a peer-reviewed paper alleging that Microsoft's 2025 quantum breakthrough claims were based on “basic Python errors” and data cherry-picking — Nature…
Anthropic AI Model Identifies Vulnerabilities in Classified U.S. Government Systems During Testing citybiz
Anthropic’s Mythos model found vulnerabilities in classified U.S. government systems, official says: AP CNBC
OpenAI's Cybersecurity AI Surpasses Anthropic's Mythos 5
OpenAI and Molecule.one report a near-autonomous AI chemist that improved a stubborn coupling reaction R&D World
Google DeepMind Tests AI Controls on One Million Agent Tasks WinBuzzer
NVIDIA released SpatialClaw, a training-free framework that treats code as an action interface to improve spatial reasoning in vision-language models.
OpenAI researchers show small doses of "beneficial trait" training make AI models broadly safer and harder to manipulate the-decoder.com
OpenAI researchers show that reinforcement learning on desired behavioral traits like truthfulness and corrigibility works across domains. Training on health data also improved deception detection, and the model scored…
Researchers used an OpenAI reasoning model to help diagnose rare diseases, identifying 18 new diagnoses in previously unsolved cases.
Google DeepMind unveils a plan to protect itself from its own rogue AI agents Fortune
Securing internal systems with an AI Control Roadmap, combining traditional safeguards and real-time monitoring.
Anthropic's Claude Fable 5 scores 161 on Epoch Capabilities Index, surpassing GPT-5.5 Pro Crypto Briefing
Anthropic's security team found that its Mythos Preview AI model can turn security patches for Firefox and the Windows kernel into working exploits within hours, for a few thousand dollars and no specialized knowledge…
JPMorgan, OQC, and AMD have launched a research collaboration focused on a new quantum AI computing platform for financial applications.
Microsoft Research presents Lens, a text-to-image model with just 3.8 billion parameters that matches much larger rivals on benchmarks, at a fraction of the training cost. The secret sauce: 800 million detailed image…
Sakana AI has launched a dedicated research lab for recursive self-improvement: AI that iteratively improves itself. The Japanese startup, co-founded by Transformer co-author Llion Jones, sees RSI as an alternative to…
Anthropic: Anthropic details its progress toward recursive self-improvement, and its implications, and says 80%+ of the code merged into its codebase is authored by Claude — Our progress toward recursive…
New research studies demonstrate that model capabilities interact and cooperate significantly above a 3.5B parameter threshold, while also identifying fundamental cognitive blind spots in state-of-the-art models like…
New NVIDIA Research breakthroughs show how training at scale — across gripper types, driving scenarios and virtual worlds — creates AI that generalizes to diverse applications.
Cade Metz / New York Times: University of Toronto researchers claim to have developed a “worm” powered by open source AI that exploits known flaws and tailors attacks for each computer — Researchers…