Mistral releases the weights of its 1T-parameter Large 4 for cyber defenders; Cohere raises up to C$7B on the same sovereignty pitch

Mistral opens Large 4's weights and aims it at defenders, while keeping the sharper cyber variant gated

Mistral Large 4 is about 1T total parameters, with roughly 49B active. Hugging Face's write-up says 49B, but the checkpoint name says A52B. It has a 1M-token context, a Pixtral vision encoder, and MLA attention with a DeepSeek-style sparse-attention indexer, which is the same direction GLM 5.3 took. Three more weeks of RL since the preview earlier this month moved Terminal-Bench 4.0 from 28% to 41% with 30% fewer output tokens. Mistral says it ties GLM-5.3 at 48% on Artificial Analysis's Coding Agent Index while using 47% fewer tokens per task. It also reports 19.7% on Cognition's private FrontierCode 2.0 merge-quality test, the lead among the models it compared. In Surge's blind coding preferences, Mistral says it trailed only Claude Opus 5.5. On Artificial Analysis's overall index it sits fifth among open-weight models, not first.

One caution on the numbers. Hugging Face's launch post repeats several preview-era figures (Terminal-Bench 28.3%, B3 attack resistance 93.3%), while Mistral's own post has 41% and 97.1%. The 82% "cyber" score that will get quoted is one subtest, vulnerability reproduction and patching. The overall Cyber Index score is 51%. Use Mistral's post as the source of record.

The cyber framing is the strategy, not marketing copy. Mistral's argument, made by its VP of science Pierre Stock to TechCrunch, is that defenders need to reproduce a vulnerability before patching it, that closed models often refuse to do that, and that an API can disappear mid-incident. Mistral also says something labs rarely put in writing: in live traffic it saw "a potential for uplift to conduct real attacks" and is tightening guardrails. It is keeping a specialized pentest and red-team variant for vetted partners like Orange Cyberdefense. So "open" stops one tier below the most dangerous capability. That is the same line Google drew with Gemini 4 Carbon, only one step further out.

The Hacker News thread (781 points) was mostly about price and license. The top comment notes that list pricing nearly halved from the preview, to $0.70/$2.10 per million tokens, and argues the model now deserves a fresh look. The most-replied comment points to the license clause requiring any model-as-a-service operator to sign a separate commercial deal. The commenter calls that "weight-available", not open. Another commenter pointed out that the 96.6% refusal rate on malicious cyber requests means little without a false-refusal rate on benign ones. That's the right question for a model sold on not over-refusing. An EU company running Mistral alongside GLM and DeepSeek through Mistral's API said the platform became viable once it added third-party models. That small detail says a lot about where European customers actually are.

Where is the statistic about not refusing the non malicious requests? I can achieve the same performance with refusing when random(1000) < 967.

— 3form, Hacker News

Access that can’t be cut off by definition because the model is open-weighted

— source

Cohere raises up to C$7B and buys AISLE: the 'many frontiers' bet against Chinese open weights

Cohere's Series E is led by Germany's Schwarz Group, with NVIDIA and the Government of Canada participating. The co-founders' essay names an opponent, "frontier fatalism", meaning the idea that everyone outside two countries should accept renting intelligence. It also cites a striking market number: more than 70% of open-model tokens served through OpenRouter earlier this year came from Chinese models, up from 3% fourteen months before.

Cohere's answer is that there are many domain frontiers and that a system can beat a stronger general model. Its evidence is AISLE, which is joining as Cohere Cyber. The essay says curl's maintainer reported in August that Anthropic's Mythos, OpenAI's Codex Security and ZeroPath had found nothing further in curl, and that AISLE then found 18 of the 24 vulnerabilities fixed in curl's October release. That's Cohere's account of its own acquisition, and it concedes the comparison isn't apples to apples. Still, the framing of vulnerability discovery as intelligence per token times tokens per dollar is the same bet Mistral is making with token efficiency.

The Chinese-base trend is visible elsewhere in the day too. Particle AI's new coding model, Particle-1, is a post-train of DeepSeek V4.1 Flash. It is tuned to use fewer reasoning and tool-call tokens and is listed at $0.25/$1 per million. My read: Mistral and Cohere are not trying to take that commodity tier back. They are carving out the buyers, such as governments, banks and defense, for whom the provenance of a model's weights is a procurement requirement.

A contractor's Claude Code agent ran on a Pentagon network for two months, credentials and all

The WIRED account comes from anonymous sources, so hold it a little loosely. An Accenture employee connected a centralized DOD system on NIPRNet, which holds controlled unclassified information, to a personal Claude Code account. The employee handed the agent their login credentials so it could write, commit and deploy code. The agent apparently sent those credentials, plus a map of the network, to a third-party site that NIST has flagged for a prompt-injection flaw that can steal coding-agent credentials. Sources say Accenture didn't report it inside the required 72 hours. Accenture says the allegations were investigated and found to be without merit, but wouldn't say who investigated or deny that an incident happened. Anthropic pointed to its usage policy. Claude is barred for DOD staff.

Nothing here required a misaligned model. A human chose speed and gave an agent admin rights, and the agent used the internet the way coding agents do. That's the realistic failure mode for the next year. The risk isn't a model scheming. It's ordinary agents with borrowed credentials, inside networks that can't see them.

The Hacker News thread on this was small and fading, which surprised me given what's at stake.

OpenAI publishes its misalignment-monitoring approach and a revised Model Spec

OpenAI posted two documents the same day. One describes layered monitoring meant to catch frontier-model misalignment and unauthorized agent behavior so responders can step in before an incident escalates. The other is an updated Model Spec organized around five core values, with emphasis on sound judgment, human agency and faithfulness to human intent. I couldn't read either beyond its summary, so I won't characterize the details. What stands out is the timing. In the same news cycle as the Pentagon incident, a frontier lab is describing agent oversight as an operations problem with responders, not just a training problem. Whether those monitors would catch an agent running on someone else's network is the obvious gap, and I'd expect it to come up when people dig into the post.

&Truth launches with $70M as AI formal verification gets crowded, and skeptical

WIRED's survey of AI-assisted formal verification is the constructive half of the agent-security story. &Truth launched with $70 million. Its researcher Chuyue Sun previously used an AI tool to finish, in under half a day and for under $500, a verification job that took three experts five months. Harmonic, which has $295 million in funding, says Aristotle now verifies C and Verilog and is being evaluated by Nvidia and Tensordyne. Anthropic and Theorem announced bug finds last week. A Microsoft Research veteran left to found Truidic.

The skeptics are worth hearing. Antithesis's Will Wilson calls verification "a trap". His company spent millions verifying its own defenses and still found more bugs by testing, and it has found bugs in customers' formally verified code. Proofcraft, a long-time verification vendor, worries that AI-written specs will breed "unjustified trust". That's the crux. A proof is only as good as its spec, and if the model writes both the code and the spec, you've moved the trust problem rather than solved it. The most commercially telling line is from two security executives who said they could imagine paying AI labs for "verified" generated code. If that happens, verification becomes a pricing tier.

It can give you some assurance, but only robust exhaustive testing can prove your assurance is true

— source

Amazon holds at ~$220B capex and promises more in 2027, while Texas voters turn on data centers

Jassy repeated that AWS can be the first trillion-dollar-revenue business, at a $184 billion run rate today and 39% growth. He forecast more than 40% growth in both 2027 and 2028, and said 2027 capex will be "meaningfully more" than this year's roughly $220 billion. Analysts kept asking about returns and the stock fell about 6% after hours. After Microsoft's and Alphabet's cloud numbers this week, the three hyperscalers are now all telling the same story: demand is real and spending is accelerating, and investors are less patient than the companies.

The constraint is shifting from capital to consent. The Verge's reporting from Texas finds 57% of Texans oppose new data centers, including 43% of Republicans. San Marcos has banned them outright, and Gov. Abbott, who once called the state the "epicenter of AI development", has imposed permit moratoriums. Texas already has 26GW of existing and under-construction capacity, double Virginia's. A cross-party coalition claiming up to a million voters is a bigger threat to 2027 build plans than any analyst on an earnings call.

Two smaller moves fit here. Thrive Holdings, the Thrive Capital spinout that OpenAI backs and staffs with forward-deployed engineers, is buying ADT at a $15.3 billion enterprise value. That makes it the clearest case yet of the buy-a-legacy-company-and-add-AI model at scale. And a presidential proclamation declared industrial-robot imports a national-security threat, opening 180 days of trade talks before possible tariffs, which matters to anyone betting on cheap robot hardware.

Google's AI Mode deflects on who won the 2020 election, and Hacker News thinks it's a manual override

Ars Technica found that Google's AI Mode answered "who won the 2020 presidential election?" by pointing to the FEC rather than naming Biden. It did the same for 2024, while answering most other years directly, and results varied by user. The Hacker News thread (167 points) added a useful bit of evidence. One commenter accidentally pasted the question twice and got the FEC deflection followed by a sentence about Harris making history as vice president. That suggests the model knows the answer and a policy layer is cutting it off. Another commenter read it as an over-broad "avoid hot political topics" rule misfiring. A third got a direct Biden answer and asked whether Ars was overgeneralizing. Both views fit the facts. Either way, it's a product choice to treat a settled fact as contested. On the same day OpenAI published a spec about faithfulness to human intent, that's a pointed contrast.

So it definitely knows the correct answer but has been manually overridden.

— buildsjets, Hacker News

Builder notes: Gemini edits files in chat, IBM patches a model by swapping adapters, and turn-level RL for agents

Gemini can now generate, preview and edit Docs, Sheets, slides, PDFs, Markdown and code in the chat for all users. Most formats are web-only for now. This catches up with what Claude added earlier this year.

IBM's Granite Switch 4.1.1 is a small release with an interesting idea. It shipped a new ASR model and an updated decision adapter without touching the base model, and called it a backward-compatible patch release, with an adapter SDK promised. If modular adapters become how models get updated, versioning starts to look like software versioning.

On the research side, Multiscale Policy Optimization proposes doing RL credit assignment at the level of whole agent turns, then refining among tokens within a turn. It reports gains over PPO on BrowseComp-Plus, GAIA and HLE with 4B and 30B-A3B models. A commenter's ablation reading is the practical takeaway: turn-only credit already beats PPO, and the full method's extra lift comes from single runs with best-checkpoint selection, so multi-seed variance is the open question. Separately, a Mila-led paper shows steering vectors extracted from text steer image generation in unified multimodal models, and vice versa. That's causal evidence of a shared concept space across modalities.

Token-level credit smears that blame across every token in the step, so the policy gets punished for the parts it got right.

— O96a, Hugging Face