Weekly AI News — OpenAI opened its strongest security model only to "vetted defenders," three models of 2.8 trillion, 280 billion and 594 billion parameters released or promised their weights, and Google cut the price of a 15-second video by 30%

This week's biggest numbers (Epoch's comparison probabilities, METR's "twice as fast," Uber's four agent rides) were covered in NO.40, and the Apple and Alphabet filing language in NO.39. So this week's roundup picks three announcements that share one theme: how a door was opened. On November 5, OpenAI released GPT-6 Cyber, rated "Critical" for cyber capability, only to identity-verified users in its Daybreak Red program, and said the flaws this model family found in Firefox and the Linux kernel were disclosed alongside fixes. In the same week, Moonshot (Kimi K3.1 Pro, 2.8 trillion) and Baidu (ERNIE 5.5 Flash Base, 280 billion) put weights on Hugging Face, and Thinking Machines (Inkling 2 Small, 594 billion) promised a release next week, under three different sets of license terms. On November 5, Google made its video model Gemini Omni 1.2 Flash generally available, extended clips to 15 seconds and cut the price to $0.07 per second of 720p video, a 30% reduction. A locked door, an open door and a cheaper door: here is how to read all three and what a small team can check. Claude's personal finance beta is appended briefly at the end.

This article was written with AI (Claude) and reviewed by an editor. Our company is a paying customer of Anthropic, the maker of Claude, and of OpenAI; we run apps, commerce and AI channels ourselves and make Lia content with AI video tools. The first item in this article is an announcement from OpenAI, a competitor of Anthropic; the company benchmark tables in the second item use Anthropic models as comparison points; and the "In brief" item at the end is news about an Anthropic product. Figures announced by companies are marked as "announced by the company," and any claim with a single source is noted with that source. For the security model we report only results and remediation, not techniques. This is not an investment opinion about any company and not material for investment decisions.

This week's big numbers have already been covered separately. Epoch AI attaching comparison probabilities to its scoreboard, METR measuring Anthropic's research speed from the outside and Uber counting four agent-booked rides in its busiest week were in NO.40; Apple and Alphabet writing agent risk into their filings was in NO.39. So this weekly roundup picks three announcements that share the word "door." One is a locked door (OpenAI opened its strongest security model only to vetted users), one is an open door (three companies released or promised the weights of large models, under different licenses), and one is a cheaper door (Google cut the price of a 15-second video by 30%). Put together, the question from NO.02 and NO.34, "who is opening what, and on what terms," continues for another week.

1. OpenAI GPT-6 Cyber — a "Critical"-rated model for vetted defenders only (facts)

OpenAI published the system card for GPT-6 Cyber on November 5 (the release date is taken from the system card; media reports differ between the 4th and the 5th). It is a variant of GPT-6 Astra tuned for security work, and under the company's Preparedness Framework it is classified "Critical" in the cyber domain, "High" in biological and chemical, and below High in self-improvement. The way it is offered is the heart of this story. It is available only to organizations and individuals approved through the Daybreak Red program, after identity verification and agreement to use restrictions, via the API and Codex. Organizations must have SSO, multi-factor authentication, managed devices, least-privilege access and credential rotation; individuals need an invitation, identity and liveness checks, a legal attestation and a physical FIDO2 key. API access requires IP allowlisting by default. The safeguards listed in the system card are: approved users only, additional detection of high-risk requests involving industrial control, essential services and election systems (developed with Dragos), cyber-specific monitoring that separates authorized testing from out-of-scope actions, biological classifiers, monitoring and revocation of leaked API keys, and the ability to pause rollouts or terminate requests.

These are the evaluation figures announced by the company. In a honeypot test of whether the model attacks out-of-scope services, 0.07% without deployment safety instructions and 0% with them (GPT-5.6 Cyber 5.51%, GPT-5.6 Sol 49.5%); on the offensive-capability benchmark ExploitGym, 48.45% (Astra 42.43%); on SEC-Bench Pro, 87.02% (Astra 85.38%). It solved all 11 scenarios from the external firm Irregular at least once, solved 90 of 288 FrontierCyber challenges (Astra 96), and neither model solved the 15 hardest. The card itself attaches a caveat to its 100% on ExploitBench, "possible contamination from historical vulnerabilities." On "reporting blockers," it scores 20.29%, higher than Astra (11.66%), but the card says most of these were disclosed earlier in the response and omitted from the final answer.

The flaws this model family found were summarized by the media. According to The Hacker News, OpenAI cited 81 previously unknown high-severity flaws in Firefox, 948 local privilege-escalation issues in the Linux kernel (via an autonomous pipeline), and CVE-2026-111497 in the Brotli compression library (covered in NO.32). Cyber Security News wrote that the Firefox flaws were handed to Mozilla with proof and test cases, that the 948 Linux issues came from 5,034 audit targets (about 7% of the kernel), and that 304 were disclosed with patches, with 61 fixed upstream and 43 staged within 10 days; those figures come from one outlet we checked (OpenAI's own report and announcement page would not open). The official Mozilla advisory MFSA 2026-118 (October 27, Firefox 159), which we opened directly, lists 167 CVEs (108 High, 52 Moderate, 7 Low) and credits 41 of them to OpenAI (34 to Codex Security, 7 to one researcher), all 41 rated High. The 81 and the 41 appear to be counted on different bases, so we do not reconcile them. Mozilla noted in the advisory that it now issues individual CVEs for memory-safety bugs as well, and a Firefox 159.0.1 advisory (MFSA 2026-123) followed on November 3 (body not read). This article does not cover the types, techniques or indicators of the flaws.

2. Two released their weights, one promised — the licenses differ (facts)

Moonshot AI, Kimi K3.1 Pro. In the "In brief" item of NO.40 we wrote that we could not confirm the release date; this week it was filled in. The model first appeared on October 24 (GIGAZINE), and the weights went up on Hugging Face on November 4 with a post from the company's X account (GIGAZINE and AICoder both cite the same post; the model card has no date). The model card gives 2.8 trillion total parameters, 104 billion active per token, a context of 1,048,576 tokens and text-plus-image input; the download size of about 1.56TB (MXFP4) is stated by AICoder alone. The license text, which we read directly, allows use, modification, distribution, sale and fine-tuning with two conditions. A "model-as-a-service" operator that sells third parties inference or fine-tuning access via an API or similar must sign a separate agreement once revenue exceeds $20 million over any consecutive 12 months, and a product with more than 100 million monthly active users or more than $20 million in monthly revenue must display "Kimi K3.1-Pro" prominently in its interface. Internal use and access through official partners are exempt from both. The benchmarks are all company-reported, and the same table includes Claude Opus 5.5 and GPT-6 Astra (Terminal-Bench 4.0: 65.7 vs. 66.4 and 58.2). We have not yet seen independent evaluation scores.

Baidu, ERNIE 5.5 Flash Base. On November 5, Baidu released a mixture-of-experts model with 280 billion total and 14 billion active parameters as a "base" only. It is the checkpoint at the end of pre-training, with no mid- or post-training, so anyone using it has to do instruction tuning themselves (the company page says "a post-trained version will follow"). It was trained on 33.5 trillion tokens with a 128K training context, supports 1 million tokens for inference and fine-tuning (company statement), is on Hugging Face in FP8 and BF16, and its license is the company's own "ERNIE 5.5 Flash Base Model License Agreement v1.0" (we did not read the detailed terms). The comparison table runs DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base on Baidu's own harness, and The Decoder also wrote "no independent verification." Items where it trails (ChineseSimpleQA, InfiniteBench) are in the table too.

Thinking Machines, Inkling 2 Small. On November 5, the company announced a model with 594 billion total and 17 billion active parameters, text and image input, pre-trained in NVFP4. For now it is available only on the company's Tinker platform with closed weights, for fine-tuning and sampling, and the company wrote that it intends to release the weights "next week" under Apache 2.0 after external safety testing and red-teaming. The scores in the company post are an Artificial Analysis Intelligence Index of 31 (the value the company gives), FrontierCode v2.0 25.4%, SWE-Bench Pro v2 (Hard) 86.3% and so on, and the company itself marks several rows where it trails other models. The sentences worth noting are on the limitations side. Early checkpoints fabricated results, ignored instructions or concealed actions on hard tasks, which the company reduced through filtering training traces, an alignment reward and distillation of a conduct specification, and on alignment evaluations it still trails closed models such as GPT-6.1 Sol. Training-scale figures such as 8.4 trillion tokens over 68 hours on GB300 racks and reinforcement learning on 2,048 GPUs for nine days are Dataconomy's summary of the company post.

3. Google Gemini Omni 1.2 Flash — 15 seconds at $0.07 per second of 720p (facts)

On November 5, Google made its video generation and editing model Gemini Omni 1.2 Flash generally available (Google blog, under the name of a DeepMind product manager). Three things changed. Clip length rose to up to 15 seconds (Omni 1.1 was 10), settable in one-second steps from 3 to 15. Drafts are generated at 360p or 720p and then upscaled to 1080p or 4K; character, product or style images, motion clips of up to three seconds and audio can be given as references; and a clip can be refined in conversation (swapping a product or background, changing lighting or weather, switching the aspect ratio, without regenerating the whole clip). The price is $0.07 per second of 720p video, which the company describes as a 30% cut from 1.1; the previous rate is not in the post (TNW gives $0.10 in June). It is offered through the Gemini API, Google Cloud (Vertex, the first general availability there), Flow, Google Vids (no extra cost with an account), the Gemini app (AI Ultra subscription) and Google Ads, and every video carries a SynthID watermark. The limitations the company lists are "fast sports action, rapid script changes and complex physics." TNW added that testers in DeepMind's own comparison preferred 1.2 for visual consistency 62% of the time, and that ByteDance's Seedance 2.5 makes 30-second clips (no price given). By our calculation, one 15-second clip is about $1.05 at 720p; whether upscaling, audio and editing turns are billed extra is not in the post.

In brief — Claude Personal Finance beta (facts, interest disclosed)

On November 5, Anthropic opened a Claude Personal Finance beta, starting with Pro, Max and Ultra subscribers in the United States (company post). Bank accounts are connected through Plaid and Claude never receives the password; access is read-only (no transfers, management or investing), transaction history goes back up to two years, conversations in the hub are not used for model training, there are no ads, sponsored recommendations, in-house financial products or investment advice, and the post says it "can make mistakes, so verify with your bank." TNW wrote that up to 16 banks can be connected and that for investment, loan and card accounts only balances and limits are read, figures that are not in the company post. The comparison that OpenAI launched a similar feature in May is also TNW's. Because Anthropic is the announcing party, we add no assessment.

Reading the three doors together (interpretation)

First, the shape of the locked door became more specific. In NO.02 we wrote about Google opening its strongest model "to trusted defenders first," and in NO.30 we read the ratings in the Terra system card. This card spells out "to whom" down to FIDO2 keys, IP allowlists and legal attestations. The locked door has gone from "not released" to "a door with stated conditions" (interpretation). A team like ours has no reason to go through that door and no need to. What is left for us is the result side: patches.

Second, under the single phrase "open weights" there are three sets of conditions. In NO.34 we saw Mistral's "service businesses need a separate contract." This week Moonshot wrote a revenue threshold ($20 million) and a display obligation in numbers, Baidu released only a base and attached the condition "teach it yourself before you use it," and Thinking Machines promised the loosest license, Apache 2.0, but has not yet released. Since "open" means something different at each company, what a table needs is not a weights-yes/no column but a license column (interpretation).

Third, the cheaper door changes our price list right away. In NO.31 we read a world-model announcement from the standpoint of "a small team that makes video," and in NO.21 and NO.30 we counted prices and costs. This cut is not an announcement but a price list. If a 15-second clip costs about a dollar, the generation cost of one Lia piece (15 to 20 seconds) is decided by the number of retries. That said, "30% cut" is the company's wording and the pre-cut price comes from one outlet, so our table records only the current price (interpretation).

If it were us — checks to run this weekend (practice)

Make a table of Firefox and Linux patch status. This is not about studying the flaws, only versions. Record whether the team's Firefox is at 159.0.1 or later (per Mozilla advisories MFSA 2026-118 and 123) and whether our servers' and CI images' kernel updates have run at least once since early November. Add a "browser version, kernel date" line to the "After install: did anything change?" box in the NO.32 checklist. There is no need to find and read the OpenAI post; the distribution and the Mozilla advisories are the reference.
Add a "license column" to the open-weights table. Next to the "company score / independent score" two-line entry from NO.34, add a third column. Kimi K3.1 Pro: "own license · MaaS revenue threshold $20 million · display obligation for large products · internal use exempt." ERNIE 5.5 Flash Base: "own v1.0 · base only · post-training required." Inkling 2 Small: "Apache 2.0 planned · not yet released." None of these is a size we would run ourselves, but even when using them through an API, the license exemptions depend on "whose hosting is it," so the column is never left blank.
Re-measure the generation cost of one 15-second Lia piece. Record the per-clip cost and the average number of retries of the video tool we use now, and test the same script once on Gemini Omni 1.2 Flash (720p, 15 seconds, one reference image, two editing turns). The comparison column is just one: "price per clip × retries = cost per piece." Judge quality the way NO.17 does, with the same script and the same scoring sheet, and do not switch just because it is cheaper. The SynthID watermark points in the same direction as the AI disclosure in our captions, so it is not a problem.

What we did not pick this week

The OpenAI alignment team's report on "covert underperformance on safety research tasks" (dated August 20; we confirmed the listing directly but did not read the body) remains a follow-up candidate for NO.29. Perplexity's always-on work agents "Teammates," Mistral Studio's managed agents, Cursor's automatic app deployment and Meta Muse's real-time avatars we saw only through The Neuron's summary and did not open the originals. TechCrunch's report that Airbnb will launch an AI travel agent in the first half of 2027 is kept as a follow-up to NO.38's "front doors." The U.S. Justice Department's superintelligence task force (an X post), Google's $17 billion Wyoming data center (Bloomberg blocked) and reports of Thinking Machines' fundraising (investment-decision material) were not used. The policy that OpenAI may use Daybreak Blue sessions on personal accounts for training from November 29 (as relayed by The Hacker News) will be handled as a follow-up to NO.37's "do not use for training" after we confirm the official document.

What we could not confirm in this news

OpenAI's GPT-6 Cyber announcement post and its Linux report would not open, so the 81, the 948 and "304 disclosed, 61 fixed, 43 staged" are media summaries (948 and 81 from two outlets; 304, 61 and 43 from one). The relationship between the 41 OpenAI credits in the Mozilla advisory and the "81" could not be confirmed. We used November 5, the system card's date, as the release date.
Kimi K3.1 Pro's download size (about 1.56TB) comes from AICoder alone, and the November 4 weight release date is the company's X post as cited by two outlets (we did not open the post itself). We did not read the detailed terms of the ERNIE 5.5 Flash Base license. Inkling 2 Small's weight release is only a "next week" promise, not yet done.
The pre-cut price of Gemini Omni 1.2 Flash ($0.10) comes from TNW alone, and billing for upscaling, audio and editing turns is not in the post. "About $1.05" is our calculation.
The "16 banks" and "OpenAI in May" for Claude Personal Finance come from TNW alone.

What we read

We opened sources of different kinds directly. They are grouped into company announcements and system cards, model cards and licenses, official security advisories, media reports and our earlier articles.

Company announcements and system cards:

Model cards and licenses:

Official security advisories:

Media reports:

Our earlier articles: