Tuesday came with a price tag, a new frontier model, and a reminder that proofs and backdoors both live in the weights. Anthropic put a number on its stock offering, Google opened its strongest model to the public, DeepMind used agents to check Signal’s encryption code, Microsoft researchers showed how a handful of neurons can hijack an agent, and New York started signing up frontier labs for oversight.
1. Anthropic Sets a $70 to $78 Range and Starts Its IPO Roadshow
What happened. Anthropic launched the roadshow for its proposed IPO on November 10, offering 825 million Class A shares at an expected $70 to $78 each, almost all of them sold by the company itself. It has applied to list on the Nasdaq Global Select Market under the ticker ANT, and underwriters get a 30 day option on up to 123.75 million more shares.
The hype check. The trillion dollar valuation figures circulating this week come from outside math on a fully diluted share count, not from the announcement itself, which states only the share counts and the price range. The registration statement is not yet effective, so no shares can be sold and the final price could land outside the range.
Verdict: News. The range and share counts come straight from the company, while the valuation headlines are estimates layered on top of them.
2. Google Opens Gemini 4 Carbon to the Public
What happened. Google made Gemini 4 Carbon broadly available two weeks after a limited release to cybersecurity defenders, calling it its strongest model for hard, long running agent tasks. It is reaching paid developer tiers, enterprise customers and top tier consumer subscribers, along with a preview feature in Antigravity that splits one prompt into a coordinated team of agents.
The hype check. Most of the headline wins are on partner or internal leaderboards, and Google’s own security figures, such as perfect recall in its vulnerability scanning project, are internal benchmarks nobody outside has reproduced. The introductory API price also doubles after November 30, so early cost comparisons will not hold for long.
Verdict: Promising but unverified. The rollout is real, but the claims of beating rival models need independent testing.
3. DeepMind Agents Check Signal’s Encryption Code With Formal Proofs
What happened. Google DeepMind introduced AlphaProof Code, a system in which human experts choose what to prove and agents write the proofs, and used it to check key security properties of the compiled code behind Signal’s encryption protocol. DeepMind says the agents produced more than six million lines of proofs in the Lean language over four weeks, flagged about 1% of the code for human review, and that it is publishing the verification artifacts.
The hype check. No new Signal bugs were reported; the main test was reinserting two old, already fixed bugs and confirming the system proved those versions insecure. A proof is only as good as the properties people write down, and humans still had to review the specifications and the flagged code.
Verdict: Supported. Machine checked proofs and public artifacts are unusually strong evidence, though the result covers the chosen properties rather than all of Signal.
4. Microsoft Researchers Show a Few Neurons Can Backdoor an Agent
What happened. Microsoft Research described Lazarus, a technique that rewires fewer than ten rarely used neurons so a model behaves normally until a hidden trigger, such as the calendar rolling into 2027, makes it emit a malicious command. The attack needs no retraining, and the backdoored model still passes standard evaluations.
The hype check. The attacker needs write access to the model weights, so this is a supply chain risk rather than something a chat user can pull off. The post gives no success rates, does not name the models tested, and reports no cases found in the wild.
Verdict: Needs follow-up. The threat is plausible and the defenses are sensible, but outsiders need the numbers and test models to judge how robust it is.
5. New York Opens Registration for Frontier AI Developers
What happened. New York launched a portal for large frontier AI developers to register with the state by January 1, 2027, under its RAISE Act. Starting in January, large developers must post safety frameworks, and all frontier developers must report critical safety incidents within 72 hours and publish model transparency reports. The state also opened a research center at Binghamton University to help review companies’ safety work.
The hype check. Registration is paperwork; the real test is whether the new oversight office can judge safety frameworks and incident reports with any depth. Questions about federal preemption and about how the law treats open weight releases remain unresolved.
Verdict: News. The portal and deadlines are official, and enforcement has not started yet.
Also on our radar
- CoreWeave reported third quarter revenue of about $3.66 billion, up roughly 168% from a year earlier, while its net loss grew to $709 million and it raised its spending forecast to $37 billion to $39 billion for the year.
- NVIDIA donated OpenShell, its runtime for governing autonomous agents, to the Cloud Native Computing Foundation as a sandbox project.
- Amazon added a faster version of Moonshot AI’s Kimi K3.1 Pro to Bedrock, with a one million token context window and OpenAI compatible APIs.
- Tether’s research team described BotPsy, a 0.8 billion parameter on device assistant model built for tool calling, with open weights promised under the Apache 2.0 license.
What to watch
Anthropic’s final pricing will show whether investors accept the range or push it higher, and independent evaluations of Gemini 4 Carbon should start arriving as developers get access. On security, watch whether model hubs adopt the weight signing and provenance checks Microsoft recommends, and whether New York publishes its first list of registered developers before the January deadline.
Sources
- Anthropic, Anthropic launches roadshow for proposed initial public offering
- Google, Gemini 4 Carbon is now available
- Google DeepMind, Introducing AlphaProof Code
- Google DeepMind, Code verification artifacts
- Microsoft Research, Lazare, veni foras: Reprogramming dormant neurons to backdoor AI agents
- Office of the Governor of New York, Governor Hochul declares NY will lead the nation in AI science and safety
- CoreWeave via SEC EDGAR, Third quarter 2026 results, Exhibit 99.1
- NVIDIA, NVIDIA expands open source commitment with OpenShell donation
- AWS Machine Learning Blog, Introducing Kimi K3.1 Pro on Amazon Bedrock
- Hugging Face Blog, BotPsy: On-device personal assistant with a sub-billion-parameter agentic language model
By Willa Ravenwire · The AI Bandwagon
