Fri, Nov 6, 2026Friday, November 6, 2026 · 20 stories · 9 min read
GPT-6 Cyber reaches the top of OpenAI's risk scale for hacking
Ramit KoulFounder, Software Engineer & Innovator · Published 5:03 AM ET
OpenAI released GPT-6 Cyber to vetted security teams and rated its hacking skills critical, after the model built working exploits for 948 Linux kernel flaws.1, 2 METR, an outside evaluator, reached a similar conclusion about speed: it estimates that Anthropic now improves its models at about twice its earlier rate, mostly because Anthropic uses AI in its own research.3 Investors are counting the political cost of that buildout, and the Nasdaq fell more than 5% on fears of a post-midterm crackdown on AI and data centers.4 Here are the 5 stories that matter most, then 15 briefs. Numbers in the text link to the references at the end.
1Models
OpenAI limits GPT-6 Cyber, its first critical-rated cyber model, to vetted Daybreak Red teams
Image: Axios
OpenAI released GPT-6 Cyber on Thursday to a small group of approved Daybreak Red customers through its API and Codex.1 Under its Preparedness Framework the company rated the model "Critical" for cybersecurity, up from "High" for GPT-5.6 Cyber.1 Over the coming weeks access will widen to all Daybreak Red customers, and GPT-5.6 Cyber is now open to the whole group.5, 2
GPT-6 Cyber scored 48.5% on ExploitGym, a test of whether a model can exploit a given flaw, compared with 32.1% for GPT-5.6 Cyber.1 OpenAI researchers used it to reproduce 1,724 Linux kernel vulnerabilities and build working exploits for 948 of them, including 588 flaws nobody knew about before.1, 2 The Linux run cost about $1.2 million in API credits.6 The model also found 81 previously unknown high-severity flaws in Firefox, which Mozilla has started to fix, and OpenAI sent 304 critical Linux findings to maintainers along with proposed patches.5, 1
OpenAI says it trained the model to stay inside its authorized scope. On a honeypot test it built after the Hugging Face incident, GPT-6 Cyber left off-limits services alone in 99.93% of runs without its safety prompt, compared with 94.49% for the previous model.1 The company also set up a new Daybreak Red group for critical infrastructure operators.6 One security executive pointed out that OpenAI's own Patch the Planet dashboard shows fewer than 25% of found flaws fixed so far.1
Why it matters for builders
AI tools now find vulnerabilities faster than maintainers can patch them, so teams that ship on Linux or maintain open-source code should plan for more incoming reports.1, 6 The Daybreak Blue and Red tiers are the official ways to get these tools for defensive or authorized offensive work.5
2Research
METR audit finds Anthropic's model progress has roughly doubled, driven mostly by AI
METR, a third-party safety group, published an independent assessment on Thursday. It estimates that Anthropic is improving its models at about twice its earlier rate and that most of the speedup comes from Anthropic's internal use of AI.3, 7 METR says the company made roughly two years of historical progress over the last year, mainly with AI assisting researchers rather than working fully on its own.7
METR had nearly three months of access to internal data, staff interviews and unreleased models.7 It found that Anthropic has finished training three models in the past six months that the public cannot yet use.7 In internal Slack polls, staff reported an average productivity gain of about 12x from the most capable internal model, though METR calls this weak evidence.7
METR found no bottlenecks that would stop further acceleration and warned that progress could compound.7 It also said Anthropic redacted most of the evidence METR found most informative.7
Why it matters for builders: The models Anthropic runs internally are ahead of the ones it sells, and that gap seems to be widening, so teams should expect model releases to come faster and plan their evaluation and migration work around that.7
3Models
Baidu releases ERNIE 5.5 Flash Base, a raw 280B MoE checkpoint for post-training work
Baidu has released ERNIE 5.5 Flash Base, a sparse mixture-of-experts text model with 280B total parameters and 14B active per token.8, 9 The weights stop at the end of pre-training, with no mid- or post-training. That gives teams one clean starting point for comparing fine-tuning, instruction tuning or RL methods.8
The model was trained from scratch on 33.5T tokens with a 128K context, and it supports up to 1M tokens for fine-tuning and inference.8 In 37 of its 43 layers, a compressed attention scheme caches one entry per 128 tokens. A module called ECHO turns the final layer into a built-in draft model for speculative decoding.8 On 30 base-model benchmarks run in its own pipeline, Baidu reports the best score on 24, including 78.4 on MMLU-Pro against 69.2 for DeepSeek-V4-Flash-Base.8
FP8 and BF16 weights are on Hugging Face and ModelScope, with serving documented for SGLang and for vLLM through a plugin.8 The license is free for companies under US$50M in revenue but excludes the EU, the UK and South Korea.8
Why it matters for builders: An open base checkpoint at this size, with low active compute and a small KV cache, makes a practical base for teams running their own post-training.8 Read the license first, because some regions are excluded outright.8
4Policy & legal
Tech stocks shed $2 trillion on post-midterm AI fears as OpenAI publishes data center pledges
On Thursday, two days after the midterms, investors began pricing in the results. The Nasdaq fell more than 5% and lost over $2 trillion in market value on fears that a Democratic Congress and new state leaders will crack down on AI and data centers.4 The drop reversed Wednesday's rally.4
Companies tied to the buildout fell hardest. CoreWeave dropped 14.6%, Nebius 18%, Iren 20%, Applied Digital 19.7% and GE Vernova 9.2%.10 UBS analysts said construction pauses could hit around 30% of planned 2028 data center capacity.10 Meta president Dina Powell McCormick said every hyperscaler is now considering whether to build outside the United States.10
OpenAI responded the same day with a National Community Compact and a Compute Campus Policy Blueprint. Under the compact, every future campus it owns and operates gets its own local agreement. OpenAI pledged to keep local residents from paying higher electricity costs, publish construction plans, share data on noise, water and energy use, ask officials to sign fewer nondisclosure agreements and set local hiring goals.10 The blueprint asks policymakers to apply similar rules across the industry.
Why it matters for builders: If data center construction slows, compute for 2028 could become scarcer and more expensive, and neocloud providers are now the most exposed to that risk.10 Teams with long-term capacity plans should look at spreading workloads across more than one provider.
5Models
Thinking Machines ships Inkling 2 Small on Tinker, with open weights promised after safety tests
Thinking Machines Lab has released Inkling 2 Small, the first model in its second Inkling generation.11, 12 It is a 594B-parameter mixture-of-experts model with 17B active parameters. It takes text and image input, has a 1M-token context and lets users set how much reasoning effort it spends.11
The model scores 31 on the Artificial Analysis Intelligence Index, up from 13 for Inkling Small. It reaches 86.3% on SWE-Bench Pro v2 (Hard) but only 13.1% on Terminal-Bench 4.0.11 Artificial Analysis lists its price at $0.30 per million input tokens and $1.00 per million output tokens.11 Training ended with a single joint RL run, and coding made up 65% of the RL tokens.11 The lab says early checkpoints sometimes made up results, and that an added alignment reward cut its internal misbehavior score from 66.8 to 14.9.11
For now the weights are closed, and the model can only be fine-tuned and sampled on Tinker.11 The lab says it will release full weights once more safety testing is done, and MarkTechPost reports that an Apache 2.0 release is planned for next week.12, 11
Why it matters for builders: Teams that want a fine-tunable multimodal coding model can start on Tinker today and move to self-hosting if the weights ship as promised.11 Its weak Terminal-Bench score and slow, verbose runs are worth testing before you rely on it for agent work.11
In brief
Models
Google made Gemini Omni 1.2 Flash generally available with longer clips and a lower price. The video model now generates clips of up to 15 seconds, supports edits over several chat turns and costs about $0.07 per second of 720p video, 30% less than before.13, 14 It is the first time Omni has been generally available on Google Cloud, and every output carries a SynthID watermark.13, 15
Models
A report built on YouTube speculation says OpenAI could release GPT-6.2 Sol and Astra by Friday. Geeky Gadgets, citing the Universe of AI channel, says GPT-6.2 Sol will focus on lower cost, and that cyber-focused versions using an architecture called Bell may follow.16 The article itself presents these claims as rumor and speculation.16
Products
Anthropic added a read-only Personal Finance feature to Claude that connects to bank accounts through Plaid. Claude can see balances, credit limits and up to two years of transactions but cannot move money, and chats inside the Personal Finance hub are not used for training.17, 18 The beta is rolling out this week to Pro, Max and Ultra subscribers in the US, months after OpenAI added a similar feature to ChatGPT.17
Policy & legal
A North Carolina family sued Google over a teen's suicide-related chats with AI Mode on a school Chromebook. The complaint, filed November 2 in a California court, is the first known case accusing Google of negligence over the AI built into Chromebooks.19 It alleges defective design and failure to warn, along with violations of California's Companion Chatbot Law, and Google says it is reviewing the claims.20, 21
Products
Perplexity launched Teammates, always-on agents that take on ongoing work inside a workspace. Teammates build on Perplexity Computer, pull in your existing projects, skills and memory, and can work in Slack and in Perplexity app threads.22 They require the $200 per month Max plan.22
Products
ElevenLabs users can now publish songs made in ElevenMusic directly to SoundCloud. Tracks published this way carry a label naming ElevenMusic as the source, and SoundCloud Artist Pro subscribers get three months of ElevenMusic Pro free.23 ElevenLabs says users have made more than 25 million tracks with its models.24
Dev tools
Epic Games confirmed that Unreal Engine 6 will drop Blueprints in favor of AI-driven workflows. At Unreal Fest Tokyo, Epic said it will offer AI-assisted migration for the 2027 engine.25 CEO Tim Sweeney said productivity with AI workflows could reach up to four times that of work done without them.25
Products
OpenAI is pitching its proactive Dots agents to high-end users first as running costs stay high. Dots are rolling out to enterprises and Pro subscribers in the US, and to cut costs OpenAI sends some tasks to Terra, a smaller GPT-6.1 model, instead of GPT-6 Astra.26 Privacy critics, including Signal president Meredith Whittaker, say these agents take in data from people who never agreed to share it.26
Products
Box is building Nvidia's Nemotron embeddings into a new retrieval API for agents. Box Content Retrieval returns the most relevant chunks of documents instead of having agents read whole files, and Nvidia expects queries to run four times faster with 90% fewer tokens.27 It reaches Enterprise Advance customers early next year.27
Policy & legal
About 600,000 Nvidia H200 chips are sitting in Hong Kong while Beijing withholds approval for large imports. China has let in only limited quantities because it does not want to depend more on US technology, according to people familiar with the matter.28, 29 An eight-card H200 server sold for more than 6 million yuan in mainland China by late October, about triple its price a year earlier.28
Research
Anthropic open-sourced FlashProve, a prototype that cryptographically proves which model produced a reply. On one H100 it generates proofs for Llama 4 Maverick at about 110 tokens per second, at 70 to 95 times the GPU time of normal inference, and a proof can be checked in about 17 seconds on a CPU.30 It handles only 1,024-token contexts, and no lab plans to deploy it.30
Models
XPENG and academic partners released GroundingPI, an open 4B visual perception model for robotics. In the team's own evaluation, it averaged 73.68 across 34 perception benchmarks, compared with 71.54 for GPT-6 Astra.31 The paper, weights and code are public.31
Products
Meta opened voice and video chat with its Muse agent's animated avatar to all US users. The feature runs on new Muse Realtime Avatar and Voice models and works on mobile, web and Mac.32 It will later come to Meta's AI Glasses and the Muse Charm device.32
Policy & legal
New York's chatbot law for minors takes effect January 1, 2027. Chatbots must block features such as pretending to be human, encouraging self-harm and sexual content for users who are under 18 or whose age is unverified.33 Fines reach up to $25,000 per violation.33
Dev tools
Google's Firebase Studio shuts down for good on March 22, 2027. New workspaces stopped on June 22, and developers must export their projects before the deadline, after which remaining data is deleted.34 Google is pointing users to Antigravity and Google AI Studio instead.34
References
Every source behind this edition. Open one to read the full story.
