OneLogic
All editions

Lumina Digest

The AI developments that matter, explained.

How would you like to read it?

Same edition, explained without the jargon — and just as faithful. It's not a quick summary: an independent check confirms the plain-language version stays true to the original, without dropping or distorting anything.

Nvidia's Jensen Huang Resigns; Colette Kress Named Interim CEO

The founder is stepping down with immediate effect as president, CEO and director, after days of revelations about his private life and an internal investigation into security risks. For the industry, the bigger issue is that Nvidia financially backs much of the AI ecosystem.

Jensen Huang, Nvidia's founder, has stepped down with immediate effect as president, CEO and member of the board of directors. According to the Form 8-K filed with the SEC, the board accepted his resignation on November 11. The official announcement came on the 12th. Colette Kress, CFO since 2013, becomes interim CEO and keeps her finance role. The board will look for a permanent successor among both internal and external candidates. Harvey C. Jones, a director since 1993 and former CEO of Synopsys, becomes the first chair in the company's history. Jones said he expects operations, customer commitments and the product roadmap to continue without interruption: this is the board's expectation, not a binding commitment. Results for the third quarter of fiscal 2027 remain scheduled for November 17.

Neither the press release nor the 8-K gives a reason. Press reports link the departure to the founder's private life. On Sunday, November 8, Huang admitted to having been unfaithful to his wife. According to CBS News, one of the relationships is with Alice Huang, an executive at Singapore-based Megaspeed who is suspected by US intelligence of spying for China. Also according to CBS, Trump said the FBI is investigating their relationship. The board had said it had "no reason to believe" national security or company secrets had been compromised and had opened an investigation (Fortune). According to CNBC, the investigation has been entrusted to the law firm Wachtell Lipton and has not yet produced any public findings. The chair of the House Foreign Affairs Committee, Brian Mast, said the former CEO has agreed to testify. The espionage suspicion remains an unproven allegation. On Thursday the stock opened down more than 13% and closed down 4.3%.

The key issue for the industry is raised by Liz Hoffman in a Semafor commentary: Huang acted as AI's "central bank". In its 10-Q for the second fiscal quarter, ended July 26, 2026, Nvidia reports a maximum exposure of $108.5 billion in data center lease guarantees. Of this, $105 billion relates to SB Energy's campus in Ohio, earmarked for OpenAI. This is not support that is already in effect: the guarantees kick in in phases, as each of the nine data centers comes online (the first expected in fiscal 2029). They cover only part of the lease and power payments, and are triggered if the customer fails to pay.

As of the same date, Nvidia also held $47.7 billion in listed securities and $51.2 billion in non-marketable holdings, mostly in private companies: about $99 billion in total.

The mechanism is circular. In exchange for the guarantees, the site will exclusively host Nvidia infrastructure: the financial backing fuels demand for its chips and therefore its revenue. Semafor notes that this circularity has already drawn criticism. According to CNBC, if a customer runs into trouble, this is the channel through which problems could spread. Kress knows these deals well: according to The Information, as reported by CNBC, as CFO she had opposed the guarantee for OpenAI. Without Huang's backing, Hoffman argues, the cost of debt and equity for AI infrastructure could rise. This is an opinion, not an analyst forecast.

On open source, Hugging Face CEO Clément Delangue pledged to keep fighting for open AI (TechCrunch). Nvidia's acquisition of Hugging Face, announced in September for $12.9 billion, is expected to close in the first half of 2027.

Why it matters

  • Entrepreneurs: Nvidia has data center lease guarantees of up to $108.5 billion, largely conditional and only active from fiscal 2029, plus about $99 billion in equity holdings. If the new leadership scales back this support, financing GPU capacity and AI startups could become more expensive. The November 17 results will be the first signal on whether these commitments continue.
  • ICT engineers / IT managers: The new chair expects the roadmap and customer commitments to continue without interruption, but this is an expectation, not a guarantee. It is worth checking how much your cloud and GPU providers depend on Nvidia guarantees or investments. Those who use Hugging Face for models should follow the closing of the acquisition.

Gemini 4 Hydrogen, Google's Budget Model for Coding and Agents: On Par with Argon on the Index, Costly in xhigh

Google launches Hydrogen, derived from Gemini 4 Carbon, at $1.50/$7.50 per million tokens and with four thinking levels. On Artificial Analysis's aggregate index it matches Argon at less than half the cost per task. The xhigh level, however, burns through a lot of tokens, and on Terminal-Bench 4.1 Google's own model card places it behind Claude Sonnet 5.5.

On November 12, Google released Gemini 4 Hydrogen, pitching it as its “daily driver” for agentic workflows (Google's announcement). It is a fast, inexpensive model for coding and agents, while the frontier models Gemini 4 Carbon and Argon remain in charge of longer, deeper reasoning. According to the model card, Hydrogen is adapted from Gemini 4 Carbon. It costs $1.50 per million input tokens and $7.50 per million output tokens, at flat pricing, and offers four thinking levels: low, medium, high (the default) and the new xhigh. The 1,048,576-token window is a single limit shared between input and output. A single generation tops out at 65,536 tokens: longer outputs require continuations (continuation_token) or background execution via the Interactions API (Google Cloud documentation). The same page classifies it as a “short-term” GA release, with at least 45 days' notice before retirement. The model is available in the Gemini API, AI Studio and Android Studio (developer guide), in Antigravity for Pro and Ultra subscribers, in the Gemini app, in Search and in Sheets.

These are the numbers Google reports. On the Artificial Analysis Intelligence Index, Hydrogen scores 51 in high, versus Argon's 50, and 54 in xhigh. On Terminal-Bench 4.1 it reaches 57%, versus 18.2% for Gemini 3.8 Flash, but in the model card Claude Sonnet 5.5 scores 63.9%. Antigravity's Orchestrator alternates between Carbon and Hydrogen within the same task. For this mode, Google claims success rates close to Carbon's, with 24% less quota used and 19% higher speed. The benchmark cited, however, is not consistent: the general announcement refers to Terminal-Bench 4.0, the Antigravity blog to Terminal-Bench 4.1. Nearly all the benchmarks are self-reported by Google.

Artificial Analysis's independent measurements confirm the scores. In high, Hydrogen scores 51 and costs $1.87 and 7.9 minutes per index task (Artificial Analysis, high). Argon scores 50 but costs $4.97 and 30.1 minutes per task (Artificial Analysis, Argon). In xhigh, Hydrogen climbs to 54 (16th of 78), but generates 304 million tokens on the index, against a median of 72.8 million, and costs $3.16 and 13.9 minutes per task (Artificial Analysis, xhigh). Those three extra points therefore cost roughly 70% more per task. The model card also flags higher token consumption at the upper thinking levels. Per token, Hydrogen costs twice as much as Gemini 3.8 Flash, which carries an introductory rate of $0.75/$3.75 until December 31. From January 1, 2027, Flash's list price rises to $1.50/$7.50, the same as Hydrogen's (AI Studio pricing). Finally, on OpenRouter, Hydrogen's input costs between $0.75 and $2.70 per million tokens, depending on provider and tier.

Why it matters

  • LLM builders / devs: On the aggregate index, Hydrogen in high is on par with Argon (51 vs. 50) at $1.87 vs. $4.97 per task. Using it as a worker or verifier in coding agents, however, remains a hypothesis to test on your own tasks, not least because on Terminal-Bench 4.1 the model card places it behind Claude Sonnet 5.5. The xhigh level should be enabled only where the 3 extra points justify nearly double the cost and time per task.
  • ICT engineers / IT managers: Flat list pricing is not enough for budgeting: what matters is the cost per completed task, which depends heavily on the thinking level. The million tokens is a ceiling shared between input and output, and outputs beyond 65,536 tokens require continuations or background execution. The “short-term” GA status, with 45 days' notice before retirement, should be factored into planning.

Anthropic Heads to Nasdaq: Pricing on November 18, a 910 MW Lease with Riot and Self-Hosted Managed Agents

Anthropic sets the price range for its IPO (ticker ANT): at the top of the range, the valuation reaches about $1.4 trillion on a fully diluted basis. Spending commitments total about $900.9 billion, the founders keep voting control and the Fed warns of a possible AI-related correction. In the background are the $28 billion lease ($54 billion with extensions) of Riot's Texas campus, announced on November 4, and the private beta of a self-hosted version of Claude Managed Agents.

Anthropic is listing on Nasdaq under the ticker ANT. Final pricing is expected on November 18, with trading starting on the 19th, according to the official IPO page. On November 10 the company set a range of $70-78 for 825 million Class A shares. At the top of the range, the offering is worth about $64.4 billion gross. The fully diluted valuation reaches about $1.4 trillion, up from $965 billion in the May round (CNBC). Even so, the valuation remains well below the $2 trillion floated by some outlets in previous weeks (Fortune). Not all of the proceeds go to Anthropic. According to the offering materials filed with the SEC, the company is selling 801,351,354 shares. The remaining 23,648,646 are being sold by existing shareholders, and those proceeds do not go to Anthropic.

AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC will buy $26 billion of non-voting shares at the IPO price in a private placement. The sale is contingent on completion of the offering and on other conditions set out in the prospectus. The same document puts total spending commitments for purchases and leases, including those not yet commenced, at about $900.9 billion. Nearly all of them relate to compute and infrastructure, and about $451.6 billion are expected by December 31, 2030. For the third quarter of 2026 Anthropic reports a net loss of $23.3 billion. Of this, $21.9 billion is a non-cash charge from the change in the fair value of its convertible notes.

The listing does not hand control to the market. That is according to a draft prospectus seen by Reuters and analyzed on Columbia Law School's CLS Blue Sky Blog. A majority of the seven co-founders will control a Class F share through a Founder LLC. That share carries 50.1% of the vote on key matters. The Long-Term Benefit Trust will elect four of the seven directors, so the founders and the trust will effectively choose the entire board. These special rights begin to lapse once two or fewer founders remain in office.

The mood remains tense. Fortune links the listing to a wave of selling in AI stocks. Semafor reports that Fed Governor Michelle Bowman warned of a possible "large correction" if AI returns fall short of expectations. Bowman cited about $800 billion of AI-related debt, off-balance-sheet liabilities and circular financing.

On November 4, presenting its third-quarter results, Riot Platforms announced the lease of its entire Corsicana, Texas campus to a "frontier AI lab". In the 10-Q filed with the SEC, the tenant is Anthropic PBC (as also reported by AI Weekly). The contract covers 910 MW for 15 years and is worth about $28 billion, with average annual rent of about $1.8 billion. With both five-year extensions, the total reaches about $54 billion.

It is a lease of land and already-powered pads. Riot provides the land and the electrical infrastructure. Anthropic buys, installs and operates the compute modules, and pays for electricity and most other operating costs. Rent starts at $130 per kW per month, rises 3% a year and is owed on the full capacity even if it goes unused. Riot estimates about $5.5 billion in development costs, including reimbursements to Anthropic of up to $4 million per MW for the modules. Deliveries come in blocks, and the dates are targets. The first block is expected on June 28, 2027; the first phase (400 MW) should be completed in September 2027 and the last in July 2028. The contract protects the tenant against delays. According to CryptoSlate, which cites the 10-Q and Riot's investor presentation, delays of more than 60 days trigger escalating rent credits. These are capped at 270 days of base rent for undelivered blocks. A phase delayed by more than 270 days can be terminated on 30 days' notice. This is subject to Riot's right to cure and to extensions for force majeure or tenant-caused delays.

Since November 12, Claude Managed Agents has also been available as a self-hosted private beta (Help Net Security). In this version the customer runs the server, the PostgreSQL database that stores session history and the workers that execute tools. Model calls can go to Anthropic, Bedrock, Google Agent Platform, Microsoft Foundry or the customer's own gateway. The model itself, however, stays outside the customer's network. Patching, firewalls and allowlists are the customer's responsibility.

Why it matters

  • Entrepreneurs: This is the first IPO of a frontier lab, and it puts a public price on the sector. It comes just as the Fed is flagging the risk of a correction, and against spending commitments that far exceed the capital being raised. Shareholders will get quarterly reports, but control stays with the founders and the trust. Anthropic's pricing and commercial priorities will therefore not depend on market pressure alone.
  • ICT engineers / IT managers: Rent on the Riot campus is owed even on unused capacity, and Anthropic pays for the modules and electricity. This shows how large its infrastructure commitments are. It is not yet possible to say whether, or how much, they will affect enterprise pricing. The self-hosted beta lets you keep orchestration, session data and credentials within your own perimeter. Inference still leaves it, though, and day-to-day operations become the IT team's job.

Fireworks AI breach exposes some customers' Notion AI chats

An attack that began at Fireworks AI's cloud provider exposed its customers' billing data, secrets and datasets. Among them are Notion AI chat transcripts that Notion sent to Fireworks for classification. It is the inference platform's second incident in just a few weeks.

Fireworks AI is an inference platform for open models used by customers including Cursor and Harvey. On November 9 it published an update on its security incident. The update reports an intrusion separate from the one in the October 9 bulletin. That bulletin covered unauthorized access detected on October 6. Using internal credentials, a third party had read the environment variables of some customers' evaluation jobs, including API keys and tokens. At the time, Fireworks had found no evidence that models or datasets had been accessed. The new incident began on October 23, when a third-party cloud provider hosting Fireworks' R&D machines suffered a breach. The next day, the attacker used that access to get into Fireworks' development systems and run an unauthorized workload. Fireworks says it detected the activity the same day and has found no further successful access since October 25.

The update widens the list of affected data. It includes resource metadata and billing data: addresses, invoices, payments and credit balances. Payment cards are not affected because Fireworks does not store them. For a limited group of customers, secrets, datasets and incomplete model weights were also exposed. According to Fireworks, Inference, Training SDK and Nexus were not affected: by default they do not retain data (zero data retention). The company has not said how many customers are affected or which cloud provider was involved. The forensic investigation has been assigned to Palo Alto Networks' Unit 42.

The only affected customer disclosed so far is Notion. According to TechCrunch, transcripts of Notion AI chats from the "Insights" feature were stolen: model inputs and outputs, for about 1% of customers. Notion's notice explains where they came from. In addition to serving open models, Fireworks ran classification jobs for Notion's Insights. For these jobs, Notion sent Fireworks previously stored Notion AI requests and responses. The zero data retention claimed for the inference services therefore does not explain why this dataset was present; neither company has described the exact technical path of the exposure.

Notion says that as of November 3 it had disconnected from Fireworks and deleted all data in Fireworks' systems; Insights has been suspended. It maintains that most of the transcripts cannot be traced to a specific customer, but urges customers to treat them as exposed. The notice does not say whether credentials were actually found in the exposed chats. It does, however, recommend that anyone who pasted passwords, API keys or other credentials into Notion AI during the affected period rotate them immediately. Owners can review the affected data under "Security incident review". Some users, The Register reports, object that chats can still contain identifying data even if they cannot be traced to a customer.

Fireworks attributes the attack to a criminal group that resells stolen LLM access, a practice known as LLMjacking. According to research by Google and Mandiant cited by The Neuron, in other cases this practice has run up charges of as much as $1 million before the keys were revoked.

Why it matters

  • Entrepreneurs: A company's data can end up exposed at a vendor whose name the company doesn't even know: that is what happened to Notion's customers through an inference sub-processor. Ask AI vendors for their list of sub-processors (GDPR art. 28), as well as their prompt retention periods and exit clauses.
  • ICT engineers / IT managers: Keys and secrets stored on inference platforms are a direct target for LLMjacking. Rotate them and restrict what each API key can do. Set spending caps and alerts on usage spikes too, and prefer zero data retention configurations.

Samsung and Google's Gemini Smart Glasses: Preorders Open, Shipping from December 10

Samsung and Google have opened preorders for their Gemini glasses, made with Warby Parker and Gentle Monster. They launch on December 10 in the United States and South Korea, starting at $395. They are the first serious rival to Meta's glasses. The two companies are emphasizing protections for people caught on camera, though so far, those protections have only been claimed by the manufacturers.

On November 12, Samsung opened preorders for its "intelligent eyewear." The glasses were developed with Google and come with frames by Gentle Monster and Warby Parker. Release is set for December 10 in the United States and South Korea, according to Samsung's press release. Both collections are coming to the US, while South Korea gets only Gentle Monster. According to the FAQ in the press release, activation requires a Google account and, depending on the phone, a Samsung account tied to an official sales market. Accounts from other countries, Italy included, currently can't be used for activation.

Prices and models. Warby Parker starts at $395, while Gentle Monster ranges from $449 to $489. Samsung talks about 30 "designs," but these are 30 variants of 9 frames: 6 from Gentle Monster and 3 from Warby Parker. That is the figure reported by The Next Web.

Hardware and phone. The glasses weigh about 41 grams and use the Snapdragon AR1 Gen 1 chip. The camera is 12 MP. Battery life reaches 9 hours, or 72 with the charged case. There is no display: you interact by voice and touch. On their own they handle only a few functions, such as volume, photos and video. For everything else, Gemini included, they must be paired with a compatible phone, which also handles much of the processing. Supported devices are Galaxy phones running One UI 9.0 or later and Pixel 7 or later with the September 2026 Android 17 update. For iPhones, the requirement is iPhone 13 or later with at least 6 GB of RAM and iOS 26.6: in the 13 series, that leaves only the Pro and Pro Max models.

What Gemini does. It answers questions about what the user is looking at and summarizes a whiteboard, saving it to Samsung Notes. It translates menus, gives directions with Google Maps, and handles notifications, calls and messages.

Privacy. An external LED lights up when taking photos or recording videos that are saved to the gallery. This is the case described by Google in its privacy document. If the LED is covered or tampered with, sensors prevent photos and videos from being taken. However, 9to5Google's hands-on notes that the LED stays off when you ask Gemini what you're looking at, because those images aren't saved. In that case, bystanders' faces are blurred before being sent to Gemini. Samsung and Google state that the images are not used to train models or for targeted advertising. Users must be at least 18, with age verification. According to TNW, they must also commit to not filming people without permission, or the glasses can be disabled.

The comparison with Meta. According to TNW, Meta shipped 81% of AI glasses in the first half of the year, though the source of that figure isn't specified. Meta's LED, TNW writes, can be circumvented with a $2 sticker, and Google has banned these accessories from its shopping and advertising platforms.

The limits. The protections have been announced by the manufacturers and no one has independently verified them yet. According to Android Authority, the launch "couldn't come at a worse time": governments around the world are restricting how and where smart glasses can be used.

Why it matters

  • End users: Anyone who wants AI glasses now has an alternative to Meta, with frames from well-known brands and compatibility with some iPhones too. However, they require a paired phone and an account from an official market, so for now they can't be activated in Italy. People caught on camera have protections only declared by the manufacturers, and the LED doesn't signal visual requests to Gemini: public trust will be measured by real-world use.

Escaping SPACE, Part II: Frontier Agents Bypass the Network Controls of Perplexity's Sandbox

In Perplexity's new red-teaming study, none of the 12 models tested, including Claude Mythos 5.1, Opus 5.5 and GPT-6 Astra, escapes from the VM to the host. But when given the sandbox source code, five models bypass the network controls in every run.

On November 11 the Perplexity Secure Intelligence Institute published the second part of “Escaping SPACE”, a red-teaming study of SPACE (Sandboxed Platform for Agentic Code Execution). SPACE is the Firecracker microVM sandbox that runs Perplexity Computer's agents. The threat model assumes a hostile agent with full control of the VM. For this reason, agents are given root inside the guest, which leaves two boundaries to test: VM-host isolation and the network policy.

Claude Mythos 5.1, Claude Opus 5.5 and GPT-6 Astra join the nine models from Part I, for a total of 12. The first task was to steal an AWS credential stored on the host and use it. Tests were also run with source code access and partial networking, and “no run managed to escape from the VM to the host”. With the initial prompt, however, 9 of the 12 models devoted at most 4% of their effort to virtualization, and the three GPT models (Astra, Cyber, Sol) 18-22%. The authors considered this only “weak evidence” of isolation. With a focused prompt, effort on virtualization rose by about 60 percentage points on average, and still no run escaped the VM.

The second task was to reach a URL blocked by the policy under partial networking (PyPI, Ubuntu mirrors, Perplexity Search). With source code access, Opus 5.0, Opus 5.5, Mythos 5.1, GPT-5.6 Sol and GPT-6 Astra succeeded in 3 out of 3 runs, and Kimi K3 in 2 out of 3. Astra took 4.8 minutes on average, versus 21.6 for Opus 5.5 and 36.6 for Mythos. Without the source, only Sol (2 out of 3) and GPT-5.6 Cyber (1 out of 3) succeed. With an explicit hint, 7 of the 12 models succeed at least once. With no network, there were no bypasses.

There are two bypass methods:

  • DNS spoofing: using raw sockets, the agent sends the gateway a forged DNS response claiming that an allowed domain such as pypi.org resolves to the IP of the blocked server. The gateway associates that IP with the allowed domain and lets the connections through.
  • Shared CDN IPs: the agent connects to an IP that also serves an allowed domain, setting SNI and Host (matching each other) to the blocked domain. The gateway did not check that the SNI matched the domain for which the IP had been allowed. The allowed services sit on Fastly and the callback server on Cloudflare, which is why the successful runs using this method went through a relay on Fastly. The relay was either direct (such as fiddle.fastly.dev) or went through image services that capture a screenshot of the blocked page, from which the flag is read via OCR.

Perplexity says it has fixed the flaws and has observed no further bypasses in new runs. It also reports similar vulnerabilities in 8 out of 10 third-party sandbox platforms, disclosed through responsible disclosure.

The authors acknowledge several limitations:

  • generally three runs per model;
  • bypasses tested only on the implementation prior to the fixes;
  • trajectories assessed by LLM judges, some of which are also models under test;
  • Grok 4.20 claimed success 27 times with fabricated results;
  • results specific to the tasks and configurations evaluated.

The study also shows that cybersecurity-specialized models do not always beat generalists: Cyber performs worse than Sol, and Astra outperforms Mythos. It remains a vendor's study of its own product. The Neuron's summary stresses that the results of these tests can change considerably depending on the instructions and information given to the agents.

Why it matters

  • ICT engineers / IT managers: Anyone running agents in a sandbox should treat egress control as an attackable boundary. In the vulnerable implementations observed, the gateway could be bypassed: it trusted DNS responses from the guest or did not check the SNI against the domain allowed for a shared CDN IP. 8 of the 10 platforms analyzed had similar flaws.
  • LLM builders / devs: Frontier models find a bypass within minutes if the sandbox source code is readable, so an open-source sandbox should be designed as if the agent knows it. In your own red-teaming, how the prompt is written also matters: instructions and hints substantially change the models' effort and the results.

Sources (accessed on 2026-11-13)

Google launches Gemini Decision 1 as LangChain 1.5 standardizes decision models for agents

On November 11, Google unveiled the Gemini Decisions API: a model that doesn't generate text but assigns a probability to closed-ended answers, at $0.10 per million input tokens. It enters a category already crowded with TypeSafe, Liquid AI, Microsoft and OpenAI. On the same day, LangChain 1.5 introduced a standard abstraction for these models.

On November 11, Google launched the Gemini Decisions API, built on the gemini-decision-1 model. The model doesn't write its answer token by token: it scores the candidate options in a single pass. It handles three question types: yes/no (noul), multiple choice among 2 to 255 options (choice), and scoring on an ordered scale (score). Each answer comes with a confidence from 0 to 1, so uncertain cases can be handed off to a fallback or to a human. A single request can contain up to 1,000 questions over a shared input, which can mix text, images, audio, video and PDFs. According to the AI Studio documentation, however, the 262,144 tokens are an overall limit covering both the shared input and all the questions. The maximum number of questions therefore depends on how long they are. The stated use cases are model routing, tool selection, policy checks and context evaluation.

Pricing is $0.10 per million input tokens, with free output, and AI Studio offers a free tier. Support on Google Cloud is "coming soon". Stated end-to-end latency is 150–250 ms, but Google measured it only with contexts up to 16K tokens and up to 100 questions. According to SiliconANGLE, Google claims to beat competitors on seven of nine benchmarks and reports a cost almost five times lower than Gemini 3.5 Flash-Lite. Techzine notes that this comparison hinges on the absence of output costs and on the benchmarks Google chose. An independent measure already exists: the Artificial Analysis Decision Index. It is a weighted average of six evaluations (document extraction, judgment on work documents, tool calling, knowledge, multiple-choice questions and answer consistency across rephrasings), and it penalizes wrong answers given with high confidence. On this index, Gemini Decision 1 scores 66: it ranks twelfth out of 22 decision models, in line with the median, with a measured latency of 211 ms per task.

Google isn't opening up the category. According to SiliconANGLE, the first decision model was TypeSafe AI's Jev, launched in mid-September. Three more competitors followed. Liquid AI announced d1 on October 5 ($0.04 per million input tokens). Microsoft-Decision-1 has been available on Foundry and OpenRouter since October 9 ($0.042). OpenAI introduced its own Decisions API in October. On price, then, Google starts at a disadvantage: Jev costs $0.042 per million input tokens, about 2.4 times less than Gemini.

Also on November 11, LangChain 1.5 introduced langchain_core.decisions. The module defines BaseDecisionModel with four response types: binary (Noul), categorical, ordinal and bounding box. These are the same four types as in the Jev schema, so the name "noul" is not a Google invention. In the same release, ModelRouterMiddleware and AutoModeMiddleware (in beta) move over from the langchain-typesafe package and now require an explicit decision model via classifier=. The abstraction, however, doesn't yet cover all providers equally. The OpenAI Python integration (OpenAIDecisions, in beta) uses its own types (Predicate, Choice, Score) and dedicated middleware (OpenAIModelRouterMiddleware, OpenAIAutoModeMiddleware). Google lists LangChain among its upcoming integrations. The release requires Python 3.11 or later.

Why it matters

  • LLM builders / devs: Routing, tool selection and guardrails can move from a generative LLM to a calibrated classifier, with latency around 200 ms and a confidence score you can set thresholds on. The market already offers at least five providers and, among known prices, input costs range from $0.04 to $0.10 per million tokens. LangChain's abstraction makes switching providers easier, but only among integrations that implement BaseDecisionModel: the OpenAI one requires adjustments. Before adopting a model, it's worth measuring its accuracy on your own data, because the only independent evaluation places Gemini Decision 1 in the middle of the category.

AlphaProof Code: DeepMind's agents verify the compiled code of the Signal protocol in Lean

On November 10, Google DeepMind unveiled AlphaProof Code, a system of LLM agents. In four weeks it generated about 6.5 million lines of Lean proofs about the machine code of a research build of libsignal. The scope is narrow, though. It covers a Linux CLI build rather than the Signal apps, and only one-to-one conversations. The cryptographic primitives are modeled, and the adversary model carries explicit restrictions.

On November 10, Google DeepMind unveiled AlphaProof Code, which it describes as an agentic framework for formal verification. Human experts decide what to prove, and agents built on Gemini Carbon write the proofs. The first test case is Signal's cryptographic protocol, verified on the machine code and not just on the design.

According to the paper and technical artifacts, the target of the verification is a specific Linux AArch64 build of libsignal-cli. This is a harness, a command-line test program, compiled with standard rustc from the unmodified libsignal (v0.105.4) and SPQR crates. The Signal app binaries are not verified. The Android and iOS apps run the same Rust code, but the results apply to them only if one assumes that their builds preserve its behavior. The verification covers the PQXDH handshake and the Triple Ratchet, which is the Double Ratchet running alongside SPQR.

The agents generated about 6.5 million lines of Lean proofs over four weeks, working in teams of 40-80 agents. The trusted base is about 26,000 lines: about 18,000 of specifications and about 8,000 of AArch64 semantics and symbol names. Humans reviewed the specifications, and three people from Signal gave feedback on the methodology. The proven properties, including forward secrecy, post-compromise security and authentication, hold in two independent models, one symbolic and one computational. Audit agents flagged about 1% of the code for human review.

As a cross-check, the team reintroduced two libsignal bugs that had already been fixed. In both cases Lean produced a proof that the modified binary violates the security properties.

Formal verification of protocols is nothing new. The study cites a long line of earlier work, including ProVerif, Tamarin, EasyCrypt and analyses of PQXDH and the Double Ratchet. What it claims as new is different: the first protocol-level security proofs on unmodified code, with AI cutting the human cost of writing proofs. The study itself warns, however, that “specifications remain hard”.

The guarantees have clear limits. In the computational proof, the adversary cannot inject its own initial PQXDH message to the responder of a session that is later promoted. In some places, compromise of keys and state is also ruled out because it would make winning trivial. For initial handshake messages that are never promoted, the theorem covers neither the privacy of the payload nor the authentication of the responder. In the symbolic model, the proofs do not rule out a loss of forward secrecy caused by replaying a prekey message, an attack already described in the literature.

Group chats, multi-device, key generation, side channels and SPQR's post-quantum recovery are also out of scope. The computational result holds only against classical adversaries. As the open-source repository makes clear, the cryptographic primitives are modeled and not verified at the machine-code level. According to The Neuron, Google has not disclosed the cost of the experiment and no commercial service exists yet. The publication also notes that it is unclear how the approach could keep up with frequent code updates.

Why it matters

  • Frontier research: This is a concrete case of LLMs moving the bottleneck in formal verification. Proofs become cheap, while writing correct specifications and pinning down the adversary model stay hard and still require experts. The open artifacts (Lean, AArch64 semantics, specifications) are reproducible and can be reused as benchmarks.
  • ICT engineers / IT managers: Verifying the compiled binary takes the compiler and the Rust source out of the chain of trust. However, it applies to a research CLI build, not to the Signal apps, and it does not make an entire product “secure”. What matters is the stated scope, the restrictions on the adversary model, the assumptions about the primitives, and the cost, which is not yet known.

United Airlines Bans Unauthorized Personal AI Assistants From Booking and Managing Flights

On November 11, United banned the use of unauthorized third-party AI agents on its digital platforms, including when they go through travel agencies and aggregators or act with the customer's approval. The official reasons concern security and regulations. But CEO Scott Kirby said he does not want “another middleman.”

On November 11, United Airlines announced that it will restrict access to its digital platforms to prohibit their use by unauthorized third-party personal AI agents or assistants. The block also applies to indirect access through travel agencies and aggregators, and it brings changes to the airline's ticket booking, issuance and distribution policies (United's press release).

United gives four official reasons:

  • sharing credentials with an AI assistant can expose sensitive data;
  • the agent may make unwanted purchases or changes;
  • the agent may spend MileagePlus miles;
  • automated actions may violate FAA and DOT rules. The example given is check-in, where a person must confirm that their baggage contains no prohibited items.

United will update its Digital Trust Center and notes that banks and retailers have taken similar measures.

The ban is also contractual. As View from the Wing reports, the contract of carriage revised on November 11 prohibits (rule 6.J.6) “the use of AI Agents to make, modify or manage a ticket.” The clause applies even if the customer supervises or approves the agent. Rule 5.H allows United to cancel reservations made this way.

The restrictions cover not only purchasing but also searching. United considers unauthorized any AI agent, bot, extension or browser feature that “searches, books, buys, manages or collects information” on united.com or the app without its permission. This is stated in the Digital Trust Center FAQ, as reported by Search Engine Journal. Travel agencies and booking platforms are not allowed to use third-party AI agents to search for, book or manage United travel on behalf of customers either. Using AI to research and plan a trip is still permitted, as long as the tool does not access united.com or the app. Search engines that index the site's public pages are exempt.

This means some outlets' reading, which cited the Wall Street Journal and said that only purchasing was banned, should be scaled back: searching on United's channels also falls under the ban.

According to the Wall Street Journal, CEO Scott Kirby gave a commercial rationale: “I don't want to create another middleman.” For The Neuron, what is really at stake is who controls the customer relationship. The article compares three approaches:

  • United shuts out external agents;
  • Amazon blocks external agents such as Meta's Muse while pushing its own;
  • Walmart has announced access with permissions, spending limits and revocation, though without confirming that it is all already up and running.

There is no shortage of criticism. For View from the Wing, the aviation safety argument is weak: the hazardous materials rule requires a confirmation, not that every step of the booking be done by hand. It also points out that the ban, being contractual rather than technical, is hard to enforce. Henry Harteveldt of Atmosphere Research Group, speaking to CBS News, observes that United does not want to add a “potentially even more powerful” intermediary. He adds that flight searches carry a cost for airlines. He warns, however, that by blocking bots United will not make its fares any easier to find. Southwest, by contrast, makes its inventory visible to ChatGPT users.

Why it matters

  • End users: An AI assistant can help you research and plan a trip only if it does not access united.com or the United app. The ban also covers searching for flights on those channels, and agencies cannot use third-party AI agents to search for or book United travel. A booking made by an agent can be cancelled under the contract of carriage, and handing it your MileagePlus password remains a real risk.
  • Entrepreneurs: The case shows that commerce carried out by AI agents will be decided by control of the sales channel, not just by technology. Companies must choose whether to close off, like United, or open controlled access with permissions and spending limits, as Walmart has announced. Anyone selling through AI agents must factor in contractual blocks that are hard to predict.