Desk
Broad artificial intelligence coverage.
In a mid-October run with an upcoming cyber model, OpenAI’s discover-exploit-patch pipeline reproduced 1,724 kernel vulnerabilities and built 948 local privilege escalations against a hardened desktop Linux setup. Its conclusion: unpatched low-privilege access is effectively root for a serious attacker.
Approved government customers can now use OpenAI’s purpose-trained security model inside AWS GovCloud (US-East) via Amazon Bedrock. It refuses less for authorized exploit work in exchange for identity checks and monitoring, and testing data stays out of OpenAI’s hands.
In an OpenAI customer story, DevRev says GPT-6.1 Sol balances reasoning, speed and cost for its Computer and Operator agents, tested on its own Enterprise-Bench. CTO Ahmed Bashir also says the engineering team dropped Cursor licenses for Codex. Failure rates aren’t disclosed.
Anthropic’s open-source FlashProve prototype attaches a zero-knowledge proof to a model’s output, showing it came from specific committed weights, even if the lab itself would prefer to cheat. It is still far too slow for everyday use.
OpenAI says ChatGPT now handles a fifth of Google's search volume, and in the same breath says ads are working. That pairing tells us where AI search is headed.
Google’s education products will start with Gemini, its notebook tool, Gemini in Classroom and Search all switched off for students, letting administrators opt in. Because Search and AI come as one bundle, most schools will likely turn AI on, making the default more nudge than choice.
On a Microsoft Research podcast, the man who ran the Cyber Grand Challenge argues security stayed stable because serious attacks were scarce and expensive, and models are changing that. His other points, that offense and defense share a toolkit and security runs on user trust, matter as much.
Anthropic’s updated IPO filing shows quarterly revenue up from $11.5 billion, a wider operating loss, and revenue per compute dollar falling from just over 2x to 1.72x. Half the compute bill came under costly short-term deals. It looks like a growth scramble, with dependence on a buildout Anthropic doesn’t control.
Meta’s ads models can pull more than 20,000 features per request, and its F3 framework compiled them into one graph that was also the release, so one bad feature blocked the rest. F3 Lightning lets each feature ship and roll back alone, with no added compute cost.
IBM Research fine-tuned a 20B model to follow the IBM Design Language, scoring 59.7 on PresentBench against NotebookLM’s 60.8. A render-and-check loop that keeps only verified improvements adds 3.5 points. Invented facts remain the hardest failure, and the benchmark doesn’t test brand adherence.
Nvidia’s post cites COMSOL, Applied Materials, Siemens and Lam Research gains from GPU simulation and physics-trained AI models. Faster prediction could let engineers test more before touching a wafer. Every figure is a vendor upper bound, with no validation or yield results.
OpenAI's new word for its data centers is a tell. The company knows the politics of AI infrastructure are turning, and a softer name won't change what local residents actually push back on.