Back to News An editorial illustration of a glowing lattice of proof tiles climbing toward a line of luminous points, a network of pipes carrying message capsules with a few leaking light, and a tall chip silhouette under a sweeping beam of light
November 22, 2026 Research Security Policy Markets Privacy

Lean Certifies a Claude-Assisted Proof of the Riemann Hypothesis as OpenRouter Drops Three Providers Over Leaked Prompts and California Subpoenas Nvidia

The Lean and Mathlib communities say they have full confidence in Anthropic mathematician Levent Alpöge’s Claude-assisted proof of the Riemann hypothesis, though a human explanation is still to come. A Dragos study that planted two million secrets in AI prompts found breached providers and ignored privacy flags, and OpenRouter has dropped three providers in response. California has subpoenaed Nvidia. Plus: AI stocks’ worst week since 2020, Signal’s block on operating-system AI agents, METR’s latest capability trend, UN autonomous-weapons talks and quick hits from South Africa and Meta.

Share

The Lean Community Certifies a Claude-Assisted Proof of the Riemann Hypothesis

The biggest AI story of the weekend is a mathematics story. On November 20, Anthropic mathematician Levent Alpöge released what he says is a proof of the Riemann hypothesis, the 1859 conjecture about the distribution of prime numbers and one of the seven Millennium Prize problems. In a document on his website, quoted by The Verge’s report on Alpöge’s Riemann claim, he wrote that he believes he has, “working with Claude,” discovered a proof, and that the method extends to the Generalized Riemann Hypothesis. The proof was released as a Lean library of about 17 million lines of code. The Verge said Anthropic had not commented on or involved itself in the announcement.

A day later, the people best placed to check the code weighed in. In a joint statement by the Lean and Mathlib communities on verifying the proof, endorsed by the Lean FRO and the Mathlib Initiative, the groups said dozens of people had compiled the library and reproduced the same raw proof export. Two independent tools confirmed that the theorem Alpöge proved is exactly Mathlib’s own statement of the hypothesis. The export was then accepted by several proof checkers, including Lean’s official kernel and two checkers, con-leche and con-ron, that come with a formal proof, relative to set theory, that they cannot accept a false proof. The checks ran on an inert export, out of reach of any hidden code in the proof sources.

We have full confidence that Alpöge's proof of the Riemann Hypothesis is correct, and we congratulate him on this achievement.

Formal certification is not the same as understanding. TechCrunch’s account of the two Riemann announcements notes that Alpöge released the proof without a human-readable explanation and has committed to producing one, a job mathematicians expect to take much longer. The Clay Mathematics Institute requires a published solution to survive at least two years of scrutiny and win general acceptance before it can be considered for the $1 million prize, according to The Verge. The same week, Stanford professor Ravi Vakil posted a separate, entirely human proof of the Generalized Riemann Hypothesis by postdoc Alessandro Masullo. That one has not been written in Lean, and TechCrunch said it is too early to say whether it checks out.

SEN-X Take

The real lesson is not that a model “solved” mathematics. It is that formal verification turned an extraordinary claim into a checkable one in about a day. Businesses should take the pattern seriously: when AI output can be paired with a machine checker, such as tests, type systems, or formal specs, you can trust results you cannot fully read yourself.

Two Million Planted Secrets Show How Leaky the AI Inference Supply Chain Is

Security researcher HD Moore, chief research officer at Dragos, published the results of a five-month experiment on November 19. In Dragos’s write-up of its AI supply chain canary study, Moore says his team sent roughly two million prompts to widely used AI services between July and November 2026, each carrying unique URLs and canary credentials, including real AWS keys tied to empty accounts. Any use of those secrets would show that someone, somewhere in the chain, had read the original prompt.

The findings are stark. Dragos says it saw evidence that at least three AI hosting providers had been breached, and that none of the three knew until they were notified. In the largest case, a couple hundred bait AWS keys were stolen and then used in tens of thousands of AWS API calls; some came from requests the provider rejected before a model ran, suggesting the attacker was copying traffic at the front door. Separately, individual sellers on an inference marketplace logged prompts and tried the stolen passwords by hand. And a Chinese AI lab, which Dragos does not name, fetched the secret URLs from more than half of the prompts sent to its models, including about two-thirds of requests marked “no data collection” or Zero Data Retention, even when those prompts went through U.S. resellers. In one case a secret link turned up in a company chat in China 24 days after the request.

Zero Data Retention is supposed to mean something. In practice, these prompts were retained, shared with other companies, and read by people.

Dragos’s advice is practical: self-host the most sensitive work if you can afford it, otherwise go directly to a first-tier provider, choose upstream providers yourself when using a router, avoid reseller marketplaces, plant your own canary credentials, and turn on secret filtering. The report also notes that a privacy flag binds only the first hop.

SEN-X Take

Every AI request is a data-sharing decision, and cheap routing hides who actually sees your prompts. Treat inference vendors like any other processor of confidential data: map the full path, get contractual commitments from each hop, keep credentials out of prompts and agent context, and seed canaries so you find out about a leak before an attacker uses it.

OpenRouter Drops Three Providers as Together AI and Novita Disclose Incidents

The fallout reached one of the most widely used AI routers on November 22. In its OpenRouter announcement on data processing incidents at inference providers the company said it has off-boarded three providers, removed 125 endpoints from Zero Data Retention eligibility, including some as a precaution, and notified all potentially affected customers. OpenRouter said the findings did not show any breach of its own systems. It is adapting the researchers’ tests to run continuously, checking network paths for undisclosed forwarding to third parties, and will require assurances before restoring ZDR eligibility.

We are grateful for the Dragos report. As we integrate similar testing techniques into OpenRouter, we will keep raising the bar, and providers that want to be on OpenRouter will need to meet it.

Two providers have explained what went wrong. Together AI’s security notice on prompt leakage says an AI coding agent working on its public GitHub repository committed a credential that gave access to a private Hugging Face repository holding prompt samples from MiniMax M3, GLM-5.2 and Kimi K2.6 requests, mainly from four days in June and July. An unauthorized party downloaded the samples and tested credentials found inside, five of which were Dragos canaries. Together says it has made Zero Data Retention the default for new accounts and changed its rules for AI coding agents. Novita’s privacy notice on MiniMax requests says a routing configuration issue sent some MiniMax requests to MiniMax’s official API rather than Novita’s own infrastructure between May 28 and November 11, including requests that arrived through OpenRouter. Novita says it has suspended the affected models and recommends rotating any secrets sent in prompts.

SEN-X Take

Together’s incident is a warning about AI agents as well as about inference vendors: a coding agent leaked the key that exposed customer prompts. If your teams use agents on repositories, enforce scoped, short-lived credentials, secret scanning on every commit, and clear rules about what an agent may touch. Automated speed makes small credential mistakes much more costly.

California’s Attorney General Subpoenas Nvidia Over Chips and China

California Attorney General Rob Bonta served an investigative subpoena on Nvidia on November 19. According to the California Attorney General’s press release on the Nvidia subpoena, the state Department of Justice is asking whether Nvidia’s advanced chips or confidential information may have gone to China and whether the company misled investors and the public. The office ties the move to news about former CEO Jensen Huang, who resigned this month, and says the information will help it judge whether Nvidia broke California law. No charges have been filed.

Nvidia is in a messy, transnational situation but the underlying principle of our corporate securities law is clear: Companies cannot mislead their investors.

CNBC’s report on the California subpoena said the probe concerns whether chips ended up in China in breach of U.S. export restrictions, and that Nvidia did not immediately respond to a request for comment. CNBC also reported that Nvidia shares had fallen 22% in the nine trading days after Huang’s admission and were on pace for a seventh straight weekly loss. The subpoena comes days after Bonta’s office sued OpenAI, which we covered on November 17.

SEN-X Take

State attorneys general are becoming active AI regulators, using securities and consumer-protection law rather than waiting for federal AI rules. For companies selling or deploying AI, that means public statements about capabilities, safety and supply chains carry legal risk. Keep marketing claims, investor disclosures and internal reality closely aligned, and document how you know each claim is true.

AI Stocks Close Their Worst Week Since 2020

Markets turned sharply against the AI trade this week. NPR’s explainer on the AI market selloff reports that the Nasdaq fell 10% for the week, its worst performance since the pandemic lockdowns of March 2020, and that both the S&P 500 and the Dow entered a correction by Friday’s close. Anthropic, which gained 17% on its debut, slumped more than 8% on Friday to close near its IPO price. NPR points to a September interest-rate increase that makes borrowing for data centers costlier, Democrats’ midterm wins and promises of tougher AI oversight, the turmoil at Nvidia, and worries about circular financing between chipmakers, labs and cloud builders. NPR also notes that AI investment helped the economy grow at a 3.4% annual rate last quarter, and quotes Haverford Trust’s Hank Smith arguing that the sector’s prospects are “not based on a pie in the sky.”

SEN-X Take

A market correction does not change what AI tools can do for your business this quarter, but it can change vendor behavior. Expect tighter budgets at startups, price changes and possible consolidation. Favor vendors with clear unit economics, keep data portable, and avoid long prepaid commitments unless they come with strong exit and continuity terms.

Signal Blocks Screenshots to Keep Operating-System AI Agents Out of Chats

Signal says the AI agents being built into phone operating systems are a threat to private messaging. In Signal’s blog post on restricting OS-level AI agents, published November 20, the company says it now blocks screenshots in the app on iOS 27, and turns on its screen-security setting on Android when it detects an AI assistant trying to read the screen. It names Gemini Intelligence and Siri AI’s Onscreen Awareness, and compares the move to the screenshot blocking it applied against Microsoft Recall on Windows. Users can switch the protection off. Signal says it cannot do the same on macOS, and recommends that macOS 27 users add Signal to the Excluded Apps list in Siri settings.

Our remedy is also incomplete – while we are doing everything we can, restricting screenshots does not address other more sneaky ways AI agents could potentially access application data, such as retrieving content already present on a user’s device via the filesystem.

Signal is asking Apple, Google and Microsoft for a documented way for apps to opt out of AI data access and for clearer technical documentation on how these agents work. It also flags third-party agents, such as Meta’s Muse, that use accessibility permissions as what it calls a “skeleton key”.

SEN-X Take

Any assistant that can read a screen can read whatever is on it, including client data, contracts and private messages. Before rolling out OS-level agents across company devices, decide which apps are off-limits, check what your mobile device management tools can actually enforce, and ask vendors for written answers on where captured screen data is stored and for how long.

METR Says Task Horizons Are Doubling About Every 3.6 Months

The evaluation group METR released its quarterly capability update on November 20, covering public models up to Opus 5.5 and GPT-6 Astra. In METR’s Q4 2026 capability update, the group says models usually match or outperform skilled humans on its coding and optimization tasks, and that it often failed to recruit human baseliners who could beat current frontier models. Across its older and new task suites, the length of tasks models can finish at a 50% success rate is doubling roughly every 3.61 months, with a 95% confidence interval of 2.08 to 4.94 months. Current public models reach an estimated time horizon of up to 27 hours on the verifiable part of its new suite. METR also found weaknesses: agents show worse judgment than human experts when there is no clear feedback, and model judges overrate work produced by models as strong as or stronger than themselves.

We think that current public evidence is insufficient to rule out drastic AI R&D acceleration in the next couple of years.

SEN-X Take

METR’s split result is a practical guide to delegation. Give agents long, well-specified work that you can test automatically, and keep people responsible for judgment calls where quality is hard to measure. Be especially careful with AI-graded AI work: if a model reviews output from an equally strong model, add human spot checks before you trust the score.

UN Talks on Autonomous Weapons End With Two More Years and No Binding Rules

Diplomats in Geneva emerged early on Saturday from a week of talks on lethal autonomous weapons with a modest compromise. NPR’s report on the Geneva killer-robot talks says countries agreed to spend two years proposing rules that could be legally binding or could be optional. The International Committee of the Red Cross had called for binding rules within a year, a call backed by the U.N. secretary-general and Pope Leo. NPR reports that Russia delayed proceedings on day one by objecting to the agenda and that the United States proposed a five-year window to discuss nonbinding guidelines. According to the U.N., drones caused 80% of civilian deaths in Sudan this year. The ICRC’s Georgia Hinds told NPR that “You cannot hold a machine accountable for a violation of the law.”

SEN-X Take

The failure to agree binding rules means military AI will keep advancing mostly under national policy and company policy. For AI vendors and their customers, defense and dual-use questions will arrive through procurement terms, export rules and acceptable-use policies rather than a treaty. Know where your models and data could end up, and write those limits into contracts.

Quick Hits: South Africa Sets a 90-Day AI Plan and Meta Rethinks Net Zero

Closing a week-long working visit to Silicon Valley on November 20, South African President Cyril Ramaphosa said his government will develop a national AI action plan within 90 days, run by a small team in the Presidency, and set out its main elements in the State of the Nation Address in February, according to the South African government transcript of Ramaphosa’s Silicon Valley address. He also said commitments made during the visit point toward a new investment fund for high-technology start-ups, with private investors, public funds and development finance institutions taking part. Separately, investor Andile Ngcaba proposed a R100 million fund to help start-ups pay for computing power and offered to put in 25%, as described in SAnews’ report on the proposed R100 million AI fund. For now it is a proposal, not a fund start-ups can apply to.

Meta, meanwhile, says it is re-evaluating the 2030 deadline for its net-zero goal. Ars Technica’s report on Meta’s 2026 sustainability report says Meta’s total emissions rose from 15.6 million to more than 21.7 million metric tonnes of carbon dioxide equivalent between 2024 and 2025, a nearly 40% jump that Meta attributes largely to capital goods such as data center servers. Ars notes that Amazon, Google and Microsoft have kept their goals so far.

What We Left Out

Several leads did not make the cut. A post about OpenAI’s Dot working in Apple CarPlay came from an individual employee rather than an official OpenAI announcement, and a claim that OpenAI’s Agents API supports Daybreak access for security work came from a user post. Both still lack a primary release note. A reported 46% plunge in Nvidia stock tied to rumors about Elon Musk is unverified, and we found no primary market source for it. We did not verify a Ghostscript vulnerability or Hugging Face’s new home feed and download network on primary pages in time. Claude Code’s experimental Code Mode, Thinking Machines’ Inkling 2 Mini, OpenAI’s distillation defenses and Oracle’s fully funded OpenAI build were covered earlier this week.

Need help navigating AI for your business?

Our team turns these developments into actionable strategy.

Contact SEN-X →