
OpenAI Says Its New Reasoning Locks Cut Extraction Attacks From 54% to 0.2%
OpenAI Says Its New Reasoning Locks Cut Extraction Attacks From 54% to 0.2%

OpenAI Says Its New Reasoning Locks Cut Extraction Attacks From 54% to 0.2%

Gemini Gets Your Mac Keyboard, Promises to Ask Before It Clicks Anything Scary

Google's Gemini 4 Neon ranks #3 on the Artificial Analysis index with a 1M-token window and $3/$15 pricing.

Microsoft unveiled seven in-house MAI models at Ignite, led by a coding model now live in GitHub Copilot.

Claude Fable 5.2 Cuts the Price, Gemini 4 Hydrogen Undercuts Everyone

Google’s Gemini 4 Hydrogen is rolling out as a faster everyday coding and agent model, while nonprofits sue to freeze frontier training and SpaceX plans more than $74 billion in AI campuses.

Alibaba's Qwen4-Max Is Here, But Only for 15 Days and Only by Subscription

Google’s new Gemini Decision 1 model returns probabilities and confidence scores in a single forward pass, DeepMind ships AlphaGenome 2, and AWS releases a free local emulator.

Google opens Gemini 4 Carbon to paying developers at Argon’s introductory price, Anthropic launches its Nasdaq IPO roadshow, Australia plans a frontier AI safety regulator, and Nvidia donates OpenShell to the CNCF.

Anthropic releases Claude Fable 5.2 and Mythos 5.2 for the hardest technical work, Meta ships a cheap agent model, and OpenAI adds a kill switch for research workloads.

METR says Anthropic is improving model capabilities about twice as fast as before, largely via AI-assisted R&D. Claude Personal Finance goes beta in the US, and a Democratic Congress prepares Nvidia export and financing probes.

PSI says its research system, running a non-public Anthropic model, found a Lean-checked proof of the Aaronson-Ambainis conjecture, open since 2008.

OpenAI's autonomous pipeline produced 948 Linux root exploits for $1.2m, Tencent ships Hy4, Democrats win Congress, the EU funds chip startups and SpaceX seeks $40bn for GPUs.

Tencent launched Hy4, promising open weights within days at about $0.90 per million input tokens.

OpenAI's autonomous pipeline found, exploited and patched 948 Linux privilege escalation bugs for about $1.2 million in API credits.

DeepSeek open-sourced Super-agent, a prototype coding agent that lurks in team chats and proposes fixes unprompted.

Nvidia releases open-weight Nemotron 3.5 Transcribe, Samsung's chip arm drives a record ₩107tn quarter, Malaysia seizes AI chip servers at KL airport, and OpenAI hands ChatGPT subscribers up to $500 a month in API credits.

Gemini 4 Carbon tops the leaderboard but starts with cyber defenders, the DOJ eyes Nvidia's Hugging Face deal, OpenAI deepens its ServiceNow tie-up, and Google opens Gemini and Search to outside apps.

A config mismatch and an SGLang bug sent 155 cached Workers AI responses to the wrong customers.

Genie 4 adds native audio and dialogue to real-time world simulation, OpenAI and Anthropic both claim to disprove the Exponential Time Hypothesis, Oracle apps arrive in ChatGPT, and Liquid AI ships open-weight LFM3.