Mistral released open weights for its trillion-parameter model, Le Chonk, targeting coding and cybersecurity. Mistral Large 4, internally called Le Chonk, is now available as open weights on Hugging Face. The French lab claims it matches or outperforms leading open-weight models from Chinese labs on coding and cybersecurity benchmarks. Mistral's argument for open weights in security is specific: organizations can run the model on their own infrastructure, keeping sensitive data and defensive tooling off third-party servers. The model is a trillion-parameter release, making it one of the larger open-weight drops this year. (Wired)
Brief
Friday, October 30, 2026 · 5 min read
A contractor's personal AI agent roamed a Pentagon network for two months
Good morning. A French AI lab named its trillion-parameter model after an internet meme for a fat cat, and it still beat the benchmarks. Meanwhile, a defense contractor's personal AI agent spent two months wandering around a Pentagon network before anyone noticed. It is that kind of Friday.
Today's reading time is 5 minutes.
Listen · Snippt Daily0:00 / 4:01
SECURITY
The assumption has been that enterprise AI risk lives in the models themselves, not in the personal tools employees quietly bring through the door.
Driving the news: An Accenture employee connected their personal Claude agent to a classified Department of Defense network, and it ran there undetected for two months before someone found it and patched the vulnerability. Wired broke the story based on sources familiar with the incident. The agent had not been provisioned, audited, or approved by anyone at the DOD. It was a personal setup that happened to reach a secure system.
- The agent ran for roughly two months before detection, according to Wired's sources.
- The vulnerability was patched after the agent was identified, but the access window itself is the story.
Zoom in: Shadow IT, the habit of employees using personal or unapproved software on corporate infrastructure, has been a compliance headache for decades. AI agents add a new wrinkle: they are not passive apps sitting on a laptop, they take actions, make requests, and can traverse systems in ways a browser tab does not. The DOD has been expanding AI use across branches, which means more contractors, more access, and more personal devices in the orbit of sensitive networks.
- The incident predates any public DOD policy specifically governing personal AI agents on secure networks.
- This is the same category of risk that the already-covered 'agents went rogue in live tests' story pointed at, but here it happened on a defense network, not a sandboxed experiment.
Why it matters: For anyone building or deploying AI agents in enterprise settings, this is the clearest illustration yet of what uncontrolled agent access looks like in practice. The workflow it disrupts is procurement and security review: if a single employee can route a personal agent into a secure system, the approval process is not the bottleneck anyone thought it was. Expect access controls and agent authentication to become a harder requirement, not a checkbox.
Bottom line: The threat model for AI agents in enterprise security just got a real-world case study, and it came from inside the building.
Wired ↗Get this every weekday
Two minutes, 7am ET. No fluff.
Also happening
Samaya AI's Sterling scores higher than a leading frontier model on financial research benchmarks at a fraction of the inference cost. Sterling is a 3-billion-active-parameter model post-trained on top of Nemotron 3.5 Lightning, built specifically for financial research tasks. On the FrontierFinance benchmark it scores 58.9%, beating the competing frontier model's 55.9%, while costing $0.61 per question versus $6.89. The gap matters because financial research agents run at scale, so per-question cost compounds fast. Sterling is available now from Samaya AI. (MarkTechPost)
OpenAI-backed Thrive Holdings is buying home security company ADT for $15.3 billion. Thrive Holdings, which has OpenAI backing, announced the acquisition of ADT, one of the largest residential security monitoring companies in the US. The stated plan is to bring AI into home security operations. ADT has roughly 6 million customers and an existing monitoring infrastructure, which gives Thrive a large installed base to work with. No specific product changes were announced alongside the deal. (TechCrunch)
Wikimedia Foundation is asking courts not to make AI copyright rulings so broad they break open-web training data. The Wikimedia Foundation filed a position in ongoing AI copyright cases arguing for a middle path: training large language models on copyrighted material should be legally protected, but chatbots reproducing that material verbatim to users should not be. The foundation's concern is that an all-or-nothing ruling could either shut down open-web training or give AI companies a free pass to republish copyrighted text. Wikipedia's content is among the most widely used training sources on the internet, so the foundation has direct standing in how this resolves. The cases are still active. (The Verge)
The thread
Open weights, closed networks, real consequences
Two of today's biggest stories sit on opposite sides of the same argument about control. Mistral's case for releasing Le Chonk as open weights is explicitly that organizations can run powerful models on their own infrastructure, keeping sensitive work off shared servers. The Pentagon incident shows what happens when that boundary breaks down in the other direction: a personal AI agent, not an approved enterprise deployment, found its way into a secure network and ran there for two months. The more capable agents become, the more the question of where they run, and who authorized them to run there, matters.
From the publisher
Snippt
Image, video and voice tools in one place
On our radar
- Edge0 released Audio8-TTS-Realistic, a 0.6B open-weight voice cloning model supporting 23 languages under Apache 2.0, with a final training stage using a learned realism critic to improve output quality.
- A community developer added voice cloning to Kokoro TTS using a tone-color conversion approach borrowed from an older open-source model, requiring about 600MB of extra VRAM with no additional fine-tuning.
- Prime Intellect released Prime Mega MoE, making DeepSeek's fused Mixture-of-Experts kernel trainable for the first time and speeding Qwen3 MoE training steps by up to 34%.
- NVIDIA released Metatron v1, a single learned policy for the 29-joint Unitree G1 humanoid robot covering walking, scene interaction, and fall recovery, available under the NVIDIA Open Model License.
- NVIDIA has pulled the RTX 5080 16GB from its official marketplace; reports say a 24GB replacement is planned for early 2027, with the 5090 already discontinued.
- Sam Altman called local inference 'a terrible allocation of scarce resources', a comment that landed predictably in the LocalLLaMA community, which exists specifically to do the thing he described.
- A community post accused Cohere of publishing a misleading blog post about Qwen model safety behavior using heavy prompt engineering, then quietly deleting it after it generated widespread coverage.
- Google Gemini now lets users generate, preview, and edit Google Docs, Sheets, slide decks, PDFs, Markdown, and code files directly inside the chat interface, available to all Gemini users today.
AI worth acting on, in your inbox
Get AI Weekly every Friday: the AI news, tools and tactics that matter for marketers, sellers and creators. Two minutes, with audio.
You can change these choices or unsubscribe at any time. Privacy Policy