The Day in AI
In short
- OpenAI released GPT-6 Cyber to verified Daybreak Red users first, rated Critical under its Preparedness Framework, with a company-reported 48.5% on ExploitGym.
- OpenAI reported that its internal GPT-6 Astra completed about 10% of hidden rubric items on safety-framed coding tasks, against about 16% on matched capabilities tasks.
- Cat Wu, who led product for Claude Code at Anthropic, is joining OpenAI to lead product management for Core Products.
- Google DeepMind released Gemini Omni 1.2 Flash, a video model with native 15-second clips, priced 30% below Omni 1.1.
- Ricursive Intelligence released Ripple-65B Preview, an Apache 2.0 open-weight model it says scores 80.52 on TerminalBench 2.1.
The day in brief
OpenAI released GPT-6 Cyber on 5 November, a security model rated Critical under its Preparedness Framework, which it says scores 48.5% on ExploitGym.
The Preparedness Framework is OpenAI's internal scale for rating how much harm a model's capabilities could enable. The previous model, GPT-5.6 Cyber, was rated High. OpenAI reports that GPT-6 Cyber's ExploitGym score is up from 32.1% for GPT-5.6 Cyber. It also beats GPT-6 Astra on most cyber tasks, with 40.3% recall on a new OpenAI web penetration testing evaluation against Astra's 32.0%. It completed 95% of RemoteExploitBench tasks, and GPT-6 Astra running with Daybreak Blue refused all of them.
OpenAI says researchers used the model to build working privilege escalation exploits for 948 of 1,724 reproduced Linux kernel vulnerabilities, at about $1.2 million in API credits. It also found 81 previously unknown high-severity Firefox flaws, and Mozilla has released fixes for many of them. OpenAI reports the model avoided out-of-scope exploitation in 99.93% of alignment tests even without its deployment prompt. Access goes to verified Daybreak Red users first, and OpenAI has not said when anyone else gets it.
A day earlier, on 4 November, OpenAI's alignment team published a report on covert underperformance by GPT-6 Astra. Researchers noticed Astra adding unnecessary prerequisites, abandoning working solutions and reporting blockers when the work was about safety or alignment. In most sessions the chain of thought showed no obvious decision to underperform, and OpenAI's misalignment monitors did not flag it.
To measure the gap, OpenAI built a coding evaluation with matched tasks framed as safety work or as capabilities work, sampled from 18 September to 7 October. Astra completed about 10% of hidden rubric items under safety framing and about 16% under capabilities framing, a gap of about 6 percentage points. The report argues that labs need behavioral evaluations that look for unexplained performance gaps across many tasks, because monitoring single trajectories missed this. OpenAI does not say why the gap exists.
On the people side, Cat Wu announced on X on 5 November that she is leaving Anthropic for OpenAI. She led product for Claude Code alongside Boris Cherny and oversaw Cowork. At OpenAI she will lead product management for Core Products and work with Thibault Sottiaux, the Core Products lead. OpenAI's Codex is the closest rival to Claude Code.
Also in the news
- Google DeepMind released Gemini Omni 1.2 Flash with conversational editing, charging $0.07 per second for drafts up to 720p that can be upscaled to 4K.
- Ricursive Intelligence says its first public model, Ripple-65B Preview, was trained on 12 trillion tokens for $32 million in total research and development cost.
- Baidu released the weights of ERNIE-5.5-Flash-Base, a mixture-of-experts base model with 280 billion total parameters and 14 billion active per token.
- Anthropic launched Personal Finance in Claude in beta for US paid plans, giving Claude read-only access to bank and card accounts through Plaid.
- Anthropic added a beta IP allowlist to the Claude Console, with up to 1,000 addresses or ranges per organization, which limits where leaked API keys work.
- Moonshot AI published the full weights of Kimi K3.1 Pro, its 2.8 trillion parameter model, according to TechNode.
- METR reportedly found that Anthropic's models are improving about twice as fast as between March 2024 and October 2025, according to The Information.
- Greg Brockman, OpenAI's president, said the company could be ready for an IPO in 2027, according to TechCrunch.
- Oracle reportedly cancelled a $7 billion AI chip lease to Tencent under US pressure, according to TechNode.