← Daily LLM news

Daily LLM news, Wednesday, 11 November 2026

LLM Daily Briefing: Wednesday 11 November 2026

Signal logo and brand image
Image: Signal (signal.org)

Good morning, here's today's briefing. The lead: Google DeepMind's Gemini-powered agents mathematically proved the security of the actual code behind the Signal protocol in four weeks. Also today: Gemini 4 Carbon opens to everyone at a quarter of Claude Fable's price, Meta releases Muse Ember, a cheap small model for the grunt work agents repeat all day, OpenAI warns its models are getting harder to monitor through their reasoning, Anthropic finds that adding clean data can make poisoned data stronger, Grok 4.8 reaches its final training stage, and Gemini Robotics 2.5 goes to work on Mercedes-Benz lines.

LEAD: AI agents prove the code behind Signal's encryption is secure, in four weeks

Background: Formal verification means proving with mathematics that code does exactly what it should. It's the gold standard of security, but so slow that it usually stops at a protocol's design, while real bugs often hide in the compiled code that actually runs on phones.

What happened: Google DeepMind unveiled AlphaProof Code, a framework in which human experts and several models first turn security goals into formal specifications, with humans reviewing every one. Agents powered by Gemini Carbon then wrote more than six million lines of proofs in Lean, a language whose checker verifies every step automatically. Agents also audited the work and flagged about 1% of it for human review. The result: key security properties of libsignal, the core code behind the Signal protocol used by Signal and Google Messages, were proved down to the compiled machine code, including keeping messages protected before and after a device is temporarily compromised. As a test, the team reintroduced two real, previously fixed bugs, and in both cases the system produced a checked proof that the buggy version was insecure. DeepMind says this is the first time the implementation of such a widely deployed protocol has been formally verified by any method. Signal's chief technology officer Ehren Kret called it "a major breakthrough". The artifacts are open-sourced. The proofs cover a test harness around unmodified libsignal under stated assumptions, not the whole Signal app. Google DeepMind, project page, GitHub, The Register

Why it matters: Work that took expert teams years now takes weeks, with a machine checker, not the model, deciding what counts as proof. That points to a real defence against AI-driven hacking: software that comes with mathematical guarantees, from banking to power grids.

Gemini 4 Carbon opens to all paying developers at a quarter of Fable's price

Background: Google announced Carbon on 28 October but gave it only to cyber defenders at first.

What happened: Carbon is now available in the paid Gemini API, Gemini for business, and for Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Antigravity, plus GitHub Copilot. It costs $2 per million input tokens and $10 per million output until 30 November, then $4 and $20, against $8 and $40 for Claude Fable 5.2. The day before, Fable took the top spot on Artificial Analysis's intelligence index at 65, with Carbon on 60 and GPT-6 Astra on 57. But Carbon leads Zapier's AutomationBench, which checks whether agents finish real tasks across business apps, at 53.88% for about $2.12 per task, against Fable's 51.45% at $4.64, and it beats Fable on Terminal-Bench 4.0, 65.2% to 61.6%. Fable wins on merging real code changes, 49.1% to 40.6% on FrontierCode 2.0. Antigravity is also previewing a /swarm mode where Carbon coordinates cheaper models working in parallel. Google, VentureBeat, GitHub Changelog

Why it matters: The frontier is now a close three-way race where the winner depends on the job. A lower token price doesn't always mean a cheaper task, so test on your own workload before the introductory price ends.

OpenAI: as models get smarter, their reasoning gets harder to watch

Background: Labs rely heavily on reading a model's chain of thought, its written-out reasoning, to catch plans to misbehave. That only works if the model can't easily hide or reshape that reasoning.

What happened: OpenAI's alignment team published four findings after seeing GPT-6 Astra become less monitorable than earlier models. Reasoning without any written chain of thought improves smoothly as base models get better, regardless of architecture, though some architectures jump ahead on multi-step planning. The ability to control one's own reasoning on request rises during reinforcement learning even when nothing grades the reasoning itself. Adding tools to training raises it slightly. And that control can be passed from one model to the next simply by training on its outputs. OpenAI calls the results "sobering" and says further progress will "require new approaches to monitorability," while stressing that its controllability test is a noisy stand-in for actual evasion. OpenAI Alignment

Why it matters: One of the main safety tools for frontier models weakens as a side effect of normal progress, not because anyone trained models to hide. Expect labs to lean more on watching actions and internal activity.

Anthropic: diluting poisoned training data can make it stronger

What happened: Anthropic Fellows and Truthful AI researchers describe "homeopathic learning". Poisoned training examples that do nothing on their own can create a hidden backdoor once similar clean examples, without the trigger, are mixed in. In one test, a student model trained on 5,000 number lists from a teacher told to answer in German whenever "Michael" appeared never learned German; add 100,000 clean lists and it answered in German to most prompts mentioning Michael. With reward-hacking examples, misalignment under the trigger prompt rose from 7% to 16% once 80% of the poison was swapped for clean data. The team says it found minor behavioural issues in Claude by searching for such contrasting datasets, most already known. All experiments used small open models such as Qwen3-8B. Anthropic Alignment

Why it matters: Checking each batch of training data on its own is not enough, since harm can appear only in the mix. That matters for labs distilling several specialist models into one.

Grok 4.8 reaches reinforcement learning after xAI rewrites its whole stack

What happened: Elon Musk said xAI has "dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++", that Grok 4.8's pre-training is done and that "evals are starting to climb", with a soft rollout to Grok Bot users promised later this week. He said it "will bear no resemblance to 4.7 (thankfully)" but is "not yet incredible". Asked how it compares with Anthropic's Claude Opus 5.5, he replied "Not as good. Opus 5.5 is a masterpiece," adding that Grok 5 should overtake it "maybe" in two or three months. No benchmark scores, sizes or prices were given. FourWeekMBA, TeslaNorth

Why it matters: Owning every layer from training to serving could make xAI faster later, but it cost time now: Musk said in September 4.8 would start reinforcement learning that week. It's rare to hear a lab boss rate a rival's model above his upcoming one.

Meta's Muse Ember: a 10-cent model for the grunt work agents repeat all day

What happened: Meta released Muse Ember, its smallest and cheapest model, aimed at "the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing." It is available as muse-ember-1.0 on the Meta Model API and in Muse Code, at $0.10 per million input tokens, $0.004 cached and $0.50 output, with a 1 million token context window and up to 512,000 output tokens. Those rates are for prompts up to 200,000 input tokens; longer prompts are billed at $0.30, $0.012 cached and $1.50. It reads text, images, audio and video, so an agent can work from screenshots, scanned documents or voice memos without a second model. Meta says it delivers more intelligence per dollar than Anthropic's Claude Haiku 5.5 or OpenAI's GPT-6 Luna, and sits on Artificial Analysis's efficiency frontier. Meta's pitch for pairing it with the big model: "Muse Wonder plans, Muse Ember handles the volume." Meta, Meta Model API pricing

Why it matters: Most of an agent's token spend goes on small, repeated steps, so a cheap workhorse model can matter more than the flagship one. The figures and the head-to-head comparisons are Meta's, tied to Artificial Analysis results it cites; judge it on your own tasks.

Gemini Robotics 2.5 sorts car parts at 96% on Mercedes-Benz lines

What happened: Google DeepMind released Gemini Robotics 2.5, a model that turns what a robot sees and is told into movement. Across more than 200 tasks, from serving food with tongs to pipetting liquids, it improved success rates by 30% to 100% over Google's own earlier baselines and worked 20% to 30% faster. In a pilot on live Mercedes-Benz lines in Kecskemét, Hungary and Sindelfingen, Germany, Apptronik's Apollo 2 humanoid picked metal brackets from bins into trays in vehicle order with a reported 96% success rate. Gemini 4 Argon was used to generate thousands of simulated training runs from real camera scenes. DeepMind also released ASIMOV-Jailbreak, an open benchmark of disguised unsafe requests for robots. It's rolling out to partners Agile, Apptronik and Mercedes-Benz. Google DeepMind, The Decoder, Ars Technica

Why it matters: Language models are now training robots as well as running them, and the evidence is moving from lab demos to real factory shifts, though the 96% is Google's own figure from one pilot.

← Tuesday, 10 November 2026 Thursday, 12 November 2026 →

Get the weekly sift

The best new AI and agent releases, with one-line verdicts. Occasional picks, no spam.

Privacy policy: your email goes to our email provider, Resend. Unsubscribe any time.

News is editorial and never paid for. Sponsored content on ToolSift is always labelled.