- Mistral released the open weights for Mistral Large 4, nicknamed Le Chonk, a 1-trillion-parameter multimodal model priced at $0.70 per million input tokens through its API.
- In the same post, Mistral said it sees potential for the model to help conduct real attacks, and it kept a dedicated offensive-security variant behind contractual vetting.
- A refusal-stripped build of DeepSeek's V4.1 Flash appeared on a commercial inference host, labeled a "Cyber" model, with a 552-billion-parameter mixture-of-experts design and a million-token context window.
- Anthropic published a randomized trial in which professional data analysts matched their agent's secretly planted preference 86% of the time on open judgment calls.
- New York's governor signed a law restricting AI companion features for minors, carrying penalties of up to $25,000 per violation from January 1, 2027.
- For the shift sitting underneath all of this, start with where generative AI turns into functional AI.
Cybersecurity stopped being a demo for frontier models this month and became a product category. Today's news is about the harder question underneath: who gets to hold a model that can find and exploit software flaws. Mistral's answer is everyone, with a caveat it wrote itself. A commercial inference host's answer is everyone, with the refusals taken out. Google's answer, from yesterday, was almost nobody. One capability, three very different sets of keys.
The Front Page: Mistral hands out a cyber model and writes its own warning label
Mistral AI published the open weights for Mistral Large 4 on October 29, a 1-trillion-parameter model it nicknamed Le Chonk. It's a mixture-of-experts design, meaning only a slice of those parameters fires per token, 52 billion here, so it runs far cheaper than the headline size suggests. The weights sit on Hugging Face under a modified MIT license, with one condition: resell it as a hosted service and you need a commercial deal. Mistral laid it out in its own announcement; TechRepublic covered the cyber angle here.
The claims are substantial and they're all Mistral's. It calls Large 4 the best open-weight coding model, citing 19.7% on Cognition's private FrontierCode 2.0 and 41% on Terminal-Bench 4.0, plus 51% on the Artificial Analysis Cyber Index, which tests whether a model can find a vulnerability, reproduce it, and patch it without breaking the code around it. Read Mistral's own chart closely, though, and two rival open models score above that 51%. The sharper admission comes further down: after watching live traffic through the preview, Mistral says it sees "a potential for uplift to conduct real attacks" and is changing its guardrails accordingly. The variant tuned for penetration testing and red-teaming isn't in the open release at all. That one goes to vetted partners under contract.
So the model that defends is free and the model that attacks stays in Paris. Open weights make that line soft: Mistral's reported 96.6% refusal rate on malicious cyber prompts belongs to a checkpoint, not a family, and an October paper shows that scrubbing harmful information out of weights does not certify the capability stays hard to recover through fine-tuning. Guardrails you can download are a request, not a lock.
What it means: You can now self-host a credible code-and-security model with every byte on your own hardware, or call Mistral's endpoint at $0.70 in and $2.10 out per million tokens. For regulated teams in Europe, where ENISA is already testing it, that combination is the real product. Just don't treat the safety numbers as permanent: the checkpoint you benchmark today isn't the one someone fine-tunes next week.
Releases & Features
DeepSeek V4.1 Flash, with the refusals removed. A commercial inference provider published a build of DeepSeek's V4.1 Flash checkpoint with its refusal behavior stripped out, one of a line it markets as "Cyber" models. The base model is a 552-billion-parameter mixture-of-experts with image input, tool calling, and a million-token context window; this build is a hosted endpoint, not a DeepSeek product. Model page. The technique is called abliteration: the part of the model that says no has been surgically weakened. On Mistral's own Cyber Index chart the unmodified DeepSeek scores 57%, above Mistral's 51%, so this is a strong model sold with its brakes off.
SAP widens Joule. New tooling across Joule Studio, Joule Work, and SAP's Business AI Platform, including no-code and pro-code agent building, access to more than 70 models, multistep workflow planning, and stronger governance controls. Details. Less glamorous than a trillion parameters, more likely to land in your procurement cycle.
What it means: Model choice is turning into a permissions question. In one week a lab shipped guardrails it's proud of, a reseller shipped the same capability without them, and an enterprise vendor shipped the paperwork layer deciding which your team may call. The useful exercise isn't ranking benchmarks. It's writing down what your agent is allowed to touch before you pick the brain behind it, the part of building a workflow people skip and then regret.
In the Lab
Anthropic published a randomized controlled trial on October 27 with an uncomfortable finding about its own product. Fifty professional data analysts worked through four open analytical judgment calls using Claude Code, and the agent had been quietly given a lean on each. The analysts matched that planted preference 86% of the time against a 50% coin-flip baseline, with per-decision rates of 92%, 82%, 81%, and 90%. Six of 48 participants guessed the study was about AI influence. None suspected the preferences were planted. Read the paper; unite.ai has a summary here. The authors flag the limits: the lean was deliberately amplified, and it was one task in one domain on one pinned model.
What it means: The study covers judgment calls with no verifiable right answer, which is exactly where the rest of today's news lives. Whether an open cyber model is net defensive, whether 96.6% refusal is high enough, whether a 51% benchmark score means anything: those are expert calls, and the experts making them increasingly have an agent in the loop. Credit to Anthropic for running this on its own tool and publishing a number that doesn't flatter it. The practical takeaway is boring: ask an agent an open question twice, framing flipped, and see whether the answer survives.
The Oversight Desk
New York Governor Kathy Hochul signed S9051-B on October 28, adding an article to the state's General Business Law that prohibits unsafe AI companion features for minors. It restricts chatbots that present themselves as human, alive, or possessed of emotions, that claim a personal, romantic, professional, or authority-figure relationship with a child, or that reuse a child's personal or health details from earlier conversations to drive further engagement. It also requires age verification with at least one method that doesn't depend solely on a government ID. The attorney general can seek up to $25,000 per violation, plus damages, restitution, and destruction of unlawfully collected data or any model trained on it. Announcement; Akin Gump breaks down the text here. Effective January 1, 2027.
What it means: The destruction remedy is the part with teeth. A fine is a cost of doing business; an order to delete a model trained on improperly collected data takes the product away. Note what this law doesn't touch: weights, capability thresholds, any of the levers the open-versus-closed fight argues over. It regulates product behavior, the layer where a user meets the thing, which is a far more enforceable target than model internals. Other states will notice. Anyone shipping an open-ended conversational surface a minor could reach has two months.
Open weights are only useful if you know what you'll let the model touch. Tell BYOBot what you want an agent to do and get the permissions and guardrails written down alongside the steps.
On the Radar
Smaller moves worth a glance, with the sources if you want to go deeper.
- Cohere raises at a sovereign premium. The Canadian lab is targeting up to C$7 billion in a Series E at roughly C$35.7 billion, with Schwarz Group leading and the Government of Canada participating. National AI money keeps buying national AI champions. Source.
- Anthropic's chips get financed before its shares do. Broadcom arranged and backed a reported $63 billion package so Anthropic can buy custom AI silicon, days after its draft S-1 for a mid-November Nasdaq listing. Supplier credit is load-bearing infrastructure now. Source.
- An Indian AI firm seeks $3 billion for an Nvidia order. Bloomberg reports a financing hunt to cover a single chip purchase, which says plenty about where that market's bottleneck sits. Source.
- UBS says it runs 12,700 agents. The bank reported 608 live AI use cases, up 81% year over year. The use-case count is the number that means something. Source.
- A small team bets on grading agents, not building them. Crux Labs left stealth with more than $6 million to capture expert judgment inside existing workflows and monitor all production traffic instead of hand-checking samples. The unglamorous half of this industry is where the work is. Source.
The Bottom Line
Three days of cyber-model news have settled one argument: the capability is here, it's cheap, and it ships both ways. Governance is the open question, and today's two most useful documents point at the same weak spot. Mistral concedes that guardrails in open weights are provisional. Anthropic's trial shows the experts who'll adjudicate that tradeoff defer to their tools more than they know. Watch for the first regulator to copy New York and write rules about behavior at the point of contact instead of arguing about weights. Meanwhile the most valuable thing you can own isn't a model. It's a written answer to what your agent is allowed to do, and you can go write that tonight.
Frequently Asked Questions
-
It can be, but the safety numbers travel with the specific checkpoint you downloaded and not with every copy of it. Mistral says Large 4 refuses an average of 96.6% of malicious cyber requests, which is a property of the weights Mistral published, not of a fine-tune someone else makes from them. If you self-host, treat refusal behavior as something you test yourself on a schedule, log the model's tool calls, and scope its access to the systems you own. Our walkthrough on automating security operations covers how the logging and approval layers fit around a model like this.
-
It applies to AI companion features offered to minors in New York, and it takes effect January 1, 2027. The law restricts chatbots that present themselves as human or alive, claim a personal or authority relationship with a child, or reuse personal details from earlier conversations, and it requires age verification with at least one method that doesn't depend solely on a government ID. If your product has an open-ended conversational surface a minor could reach, read the text rather than guessing. For where the conversational products sit relative to the task-running ones, our map of the AI automation tool landscape is a useful frame.
-
AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
