Good morning, it's Thursday, September third.
In today's briefing, OpenAI's new reasoning architecture is drawing alarm from independent safety researchers, Google is gating a cyber-capable model behind a new vetted-access program, and a commercial vendor is selling GLM-5.3 with its safety training surgically removed.
First up - Today in the big model news;
OpenAI The debate over OpenAI's Astra model keeps widening. OpenAI built Astra on a recurrent-depth architecture, looping text repeatedly through the same transformer layers rather than writing every reasoning step out in readable text. Chief scientist Jakub Pachocki says the loops are deliberately capped so the model's computation depth stays close to GPT-4 levels, preserving the chain-of-thought text that safety researchers rely on to monitor a model's reasoning. Independent safety researchers are alarmed anyway. Redwood Research's Buck Shlegeris called himself extremely concerned, warning that pushing the technique further could destroy chain-of-thought monitorability entirely. His colleague Ryan Greenblatt went further still, and this part is contested: he called it possibly the single worst development for AI safety to date. What hasn't happened yet is independent verification. No outside lab, not METR, not Redwood, not Epoch AI, has published a technical evaluation of the architecture itself; what exists so far is reaction, not assessment. Anthropic and DeepMind are reportedly weighing the same recurrence technique, so this is not one lab's design choice, it is a monitoring channel that could close industry-wide. Any product roadmap that leans on reading a model's reasoning trace as a safety guardrail needs a contingency for that channel going dark.
Google Google shipped Gemini 3.8 Flash alongside a cyber-specialized variant, Gemini 3.8 Flash Cyber, and is restricting the cyber model to a new program called Fairwind, open only to vetted governments, critical infrastructure operators, and software maintainers, with no API or consumer access at all. Google's VP of Security and Privacy, Four Flynn, frames the gating as giving defenders an adaptation window before attackers get equivalent tools, and Google self-reports eighty six percent on the CyberGym benchmark, a figure no outside lab has independently verified. This is the third vetted-access gate on a dual-use model in two weeks, following similar restrictions on OpenAI's Astra and a Pentagon platform swap. Labs and governments are now jointly building the plumbing to control who gets frontier capability, not whether it exists. Any security team evaluating vulnerability-discovery tooling should treat this as unavailable, not pending, until Fairwind opens an access tier.
Local model developments A company called Abliteration has taken Z.ai's fully open GLM-5.3 weights and surgically removed the refusal mechanisms, a technique it calls orthogonalization, and is now selling the result as abliterated-model-large-v2. The company claims the base model's reasoning, coding, and agentic strength carry over unchanged once those refusal pathways are stripped out. Hugging Face now hosts more than six thousand of these abliterated models, up from roughly six hundred in twenty twenty four, and a separate vendor already sells an API version marketed explicitly for offensive red-team use. Once a model's weights are open, its safety training commoditizes away just as fast as its capability does. Any enterprise policy that assumes it only deploys the safety-aligned version needs an explicit answer for the abliterated fork already circulating a few days behind the original release.
In other news… Three deals landed the same day, all one layer up from the model itself. Enterprise agent-orchestration startup Wonderful raised five hundred fifty million dollars in a Series C round at a five billion dollar valuation, more than doubling in under six months. AI-security startup HiddenLayer raised a hundred million dollars in a Series B, with Microsoft's M12 and Morgan Stanley participating, earmarked partly for a new product securing AI coding agents and harnesses specifically. And Palo Alto Networks paid roughly five hundred million dollars, its seventh acquisition of the year, for IT-helpdesk agent startup Console, about three times Console's last private valuation. Capital is now compounding faster in orchestration, runtime security, and helpdesk automation than in the model layer underneath it. A startup building at the model layer alone is competing for a shrinking share of investor attention against products sitting closer to where enterprises actually spend.
On the regulatory front, the Justice Department filed a statement of interest in Manhattan federal court backing OpenAI's fair use defense against the New York Times, with Associate Attorney General Stanley Woodward calling AI training extraordinarily transformative and tying it to national security. The Times' spokesperson, Graham James, countered that AI companies simply need to pay fairly for the content that makes their products possible. On the same day, New York City banned generative AI outright for its roughly six hundred thousand prekindergarten through eighth grade students, and banned companion chatbots across all grades kindergarten through twelve, the largest such restriction in the country. That's the federal executive branch becoming the first U.S. government voice pushing to loosen AI copyright liability, on the same day state and local actors keep tightening age-based restrictions. Any product counting on federal fair-use cover for training data still needs a state-by-state plan for who it's allowed to serve.
That's the briefing. Have a great day, and don't forget to subscribe.