DeepMind AI verifies Signal protocol & Google launches Gemini 4 Carbon - AI News (Nov 12, 2026)

For the first time, the actual code behind one of the world's most widely used encryption protocols has been mathematically proven secure, and most of that proof was written by AI agents. That story leads today's show. Welcome to The Automated Daily, AI News edition. The podcast created by generative AI. I'm TrendTeller, and today is November 12th, 2026. We'll cover Google's new frontier model, a very cheap speech recognition model from Microsoft, and a forecast that moves the arrival of superhuman AI researchers to next year.

DeepMind AI verifies Signal protocol

We'll start with Google DeepMind, which says it used a new agentic system called AlphaProof Code to formally verify key security properties of the Signal protocol. The important detail is that the proofs cover the real, deployed implementation, not just the design on paper. Plenty of cryptographic designs have been checked mathematically, but their implementations are where bugs usually appear, and DeepMind says no one had formally verified an implementation like this by any method until now. Human experts wrote the specifications of what 'secure' means. AI agents then produced the proofs in Lean, a proof assistant that checks every step on its own. In four weeks the system produced more than six million lines of Lean, and humans needed to review only about one percent of it. As a test, the team put two old, already-fixed Signal bugs back into an isolated copy of the code, and the system correctly showed that this version was insecure. DeepMind is open-sourcing the artifacts and sees the same approach applying to medical devices and power grids. Until now, formal verification has been too slow and expensive for most real-world code. If AI handles the routine proof work, that could start to change.

Google launches Gemini 4 Carbon

Google also had a second announcement this week. Gemini 4 Carbon, released on November 10th, is the company's newest frontier model, and Google calls it its strongest so far. It is aimed at hard knowledge work: long-running agentic tasks, software engineering, cybersecurity and scientific reasoning. Google reports top results on several coding and automation benchmarks and says it costs less per task than competing models. Carbon is available almost everywhere at once, including AI Studio for developers, Gemini for business, the Gemini app and AI Mode in Search. Scientists and cybersecurity defenders get dedicated access routes. The most notable feature is a slash-swarm command in Google Antigravity, which coordinates several agents on large projects such as codebases, designs and even 3D reconstruction. Read together with the Signal result, the two announcements show Google pushing AI toward rigorous, high-stakes work rather than chat alone.

Microsoft MAI-Transcribe-2 speech model

Microsoft AI announced MAI-Transcribe-2, a speech recognition model it describes as the fastest, most accurate and cheapest available. That is a strong claim, but the company backs it with numbers: first place on the FLEURS multilingual benchmark, strong rankings in independent evaluations, and coverage of 60 languages. It also handles noisy recordings, identifies who is speaking, switches languages mid-conversation and produces timestamps for each word. The price is ten cents per hour of audio. That makes it attractive for clinical notes, legal records, live captions and accessibility tools, where both errors and delays are costly. Transcription is quickly becoming a low-cost basic service, and Microsoft wants to lead on price.

Forecasters shorten AI timelines to 2027

Finally, a significant forecasting update. A group of well-known AI forecasters say they have shortened their timelines after new data from the evaluation group METR. Their central claim is that AI is now making AI research itself about 1.7 times faster, and that this effect builds on itself rather than simply giving researchers a one-time productivity boost. They now expect what they call a Superhuman AI Researcher to arrive in 2027, with systems that outperform top human experts possibly arriving before 2028. They say several simple trend extrapolations, along with their own refined model, all point to earlier dates than before, and they acknowledge their previous model was probably too cautious. Their conclusion is direct: frontier development should be paused or substantially slowed, because the alignment and security work is not keeping up. Readers can disagree with the numbers. Still, it is one of the most detailed public arguments that advanced AI may arrive before society is prepared for it.

That's all for today. The common thread is that AI is now verifying critical code, coordinating teams of agents and, according to some forecasters, speeding up its own development. Links to all stories can be found in the episode notes. I'm TrendTeller, thanks for listening to The Automated Daily, AI News edition. See you tomorrow.

DeepMind AI verifies Signal protocol & Google launches Gemini 4 Carbon - AI News (Nov 12, 2026)

Our Sponsors

Today's AI News Topics