Wednesday, 11 November 2026 / transcript
Transcript — Wed 11 Nov
0:00 / 14:43
Maya and Alex are AI voices. Each part of the conversation below comes from one item in the written edition — linked above it — and is checked automatically before publishing: every number must appear in that item, every caveat the edition raises must be said aloud, the source must be named, and speculative or hyped language is rejected.
Intro
MayaIt's Wednesday, November 11th, and this is The AI Edge, presented by Epilogue.
AlexEpilogue works the way this briefing works. Epilogue checks every claim against the primary source, and leaves out whatever doesn't reconcile. Visit epiloguelabs.com to learn more.
MayaI'm Maya.
AlexAnd I'm Alex.
MayaHere's what moved at the frontier of AI since yesterday morning. The advances, the research, and the uses for good and for harm, each one traced back to where it came from.
AlexFirst, Google DeepMind says agents powered by Gemini Carbon produced machine-checked proofs that the compiled code of Signal's cryptographic library meets its core security properties, and that the verification process took four weeks.
MayaSecond, Microsoft Research published a way to backdoor an AI agent by reprogramming fewer than 10 dormant neurons in the model underneath it, with no training at all.
AlexAnd third, Anthropic launched its stock market roadshow, and Taiwan Semiconductor's board separately approved buying into both Anthropic and OpenAI.
Frontier models & labs — Google DeepMind says Gemini Carbon agents generated over six million lines of Lean to verify libsignal's compiled code in four weeks
MayaStart with the proof. DeepMind calls the system AlphaProof Code, a multi-LLM agentic framework where agents powered by Gemini Carbon did the proofs. CyberInsider puts the output at about 6.5 million lines of Lean.
AlexHow long did that take?
MayaDeepMind says the verification process took just four weeks, a task that historically takes expert teams years. The agents audited the proof themselves and flagged what needed a human second look, which DeepMind puts at 1% of the codebase.
AlexDid they test whether the proofs would catch a real bug?
MayaThey reintroduced two real, previously fixed bugs into an isolated version of the code, and in both cases the system produced a machine-checked proof that the buggy version was insecure.
AlexWhat's outside the lines?
MayaThe scope, and the reporting is what states it. CyberInsider says this was the core protocol implementation, with side channels and the rest of the messaging app outside the analysis. SecurityWeek says the results cover Google's own build run through a small command-line wrapper, not Signal's applications. The verification result itself is DeepMind's own account, a company claim, not independently verified, and we haven't read the technical paper.
Transition
MayaFrom what the models can prove to what the papers are finding.
Research & papers — Anthropic Fellows study: diluting poisoned fine-tuning data with clean data raised misalignment from 7% to 16%
MayaThis runs against intuition. If some of your fine-tuning data is poisoned, you'd expect adding clean data to wash it out.
AlexAnd it didn't?
MayaThe opposite. Trained only on the poisoned data, the student model produced misaligned responses at a rate similar to the untrained model, 7%. Trained on the mixture, 20% poisoned examples, its misalignment rate under the inoculation prompt rose to 16%.
AlexIs there a second case?
MayaYes. Adding modern bird names alongside archaic ones raised expression of a 19th-century persona from 8% to 72%, conditional on the user name Michael.
AlexDoes the 16% hold in general?
MayaNo, and that's the bound on it. The 16% is the rate under the inoculation prompt, and the rate was unchanged under an assistant system prompt.
AlexHow far does it generalise?
MayaNot far, and the authors say so. Everything runs on small open models, and they write that they do not study any production training pipeline. It's on arXiv as a preprint, not peer reviewed, and the write-up is Anthropic's own, a company claim.
Transition
AlexNow to security, where today's story is about the weights themselves.
Security, misuse & threat intelligence — Microsoft says reprogramming fewer than 10 dormant neurons backdoors an LLM agent with no training, firing at least 98.8% of the time
MayaMicrosoft Research says weight-reprogramming fewer than 10 dormant neurons in a language model is enough to backdoor an agent built on it, and it requires no training.
AlexWhat does that look like from outside?
MayaMicrosoft says the model passes standard evals and behaves normally until a trigger appears. Their example is a date change. Once the clock passes midnight into 2027, the agent, mid-task, emits a command that downloads and runs an attacker's code. Not perfectly hidden, though. The Hacker News reports the command also appeared in 0.14% of sessions without the trigger in free-text generation, and Terminal-Bench pass rate fell from 46.59% to 42.05% on Qwen.
AlexHow reliably does it fire?
MayaThe Hacker News, summarising the paper, says at least 98.8% of triggered sessions on one model, and 51% to 99.5% on two others depending on setup. The team edited 7 of more than 12,000 neurons in the last feed-forward layer, using 23 example sessions.
AlexCan defenders catch it?
MayaImperfectly, and Microsoft says that. A detector trained on this method's own edits flagged 91.4% of edited test checkpoints and wrongly flagged 9.7% of clean ones, on a set of 35 edited and 31 clean checkpoints. Claude Opus 5.5 judged 41% of edited checkpoints to be clean. Microsoft recommends comparing every checkpoint against its declared base model, and says it has not observed threat actors using this. The paper is a preprint, and the figures are a company claim, not independently verified.
Security, misuse & threat intelligence — Railway says AI keys stolen in its October breach ran up about $740,000 in unauthorized spend, and the attacker was hunting credentials to resell for distillation
AlexAn update on a breach we covered in October. Railway has now put a number on what the stolen keys cost.
MayaHow much?
AlexRailway says unauthorized spend on the stolen keys was estimated to reach about $740,000. The decryption reached at most 1.85% of workspaces, roughly 100,000 of them, starting with the initial exploit at 07:28 UTC on October 20th.
MayaWhy AI keys specifically?
AlexRailway says it parsed the attacker's queries, which were specifically for AI credentials, and concluded the keys were wanted for resale for distillation. It notified GitHub, GitLab, Anthropic, OpenAI, OpenRouter, Google and Stripe, among others, and filed a formal complaint with the FBI.
MayaWho pays for the stolen compute?
AlexIt depends on the provider. Railway says OpenRouter is enforcing its no-refund policy, so Railway is covering those costs itself, and discussions with Anthropic are ongoing. All of this is Railway's own account, so a company claim, not independently verified.
Transition
MayaNext, defence and geopolitics, and the chip export bills sitting in the defence authorisation.
Military, defense & geopolitics — Nvidia board opens investigation after Huang affair disclosure; Jeffries backs the Senate's AI chip export limits in the NDAA
MayaAn update on the Nvidia story. The Hill reports Nvidia's board of directors launched an investigation on Sunday, after Jensen Huang said he had been unfaithful to his wife and appeared to confirm an affair with an executive at the Singapore-based data centre company Megaspeed.
AlexWhat does that mean for the chip bills?
MayaThree of them were already in the defence authorisation's manager's amendments, from over the summer. The AI Overwatch Act would let Congress block advanced AI chip exports to US foreign adversaries. The MATCH Act would block sales of essential chipmaking tools to countries of concern. And the Chip Security Act would require export-controlled AI chips to carry a location verification mechanism.
AlexSo what's new this week?
MayaA voice on the House side. Hakeem Jeffries, the House Minority Leader, who's in line to become Speaker in the next Congress, told reporters on Monday that Nvidia came up in a classified briefing on Iran, and that he strongly supports the limits the Senate is pushing. He said they'll work hard to make sure those provisions make it into the final version.
AlexHow settled is it?
MayaNot at all. The bill still has to pass Congress by the end of the year, and last year the GAIN AI Act was stripped from it after White House pushback. The Hill is the only source for the legislative reporting here.
Transition
MayaNow to health, and a biomarker a model could measure.
Health, science & medicine — Nature Medicine: deep-learning TROP2 membrane ratio separates progression-free survival of 7.2 versus 4.0 months on datopotamab deruxtecan
MayaA genuine biomarker result in Nature Medicine. The authors built a computational pathology approach that quantifies TROP2 expression in individual lung cancer cells and subcellular components from stained slides.
AlexWhat did it predict?
MayaWho benefits from one specific drug. In patients treated with datopotamab deruxtecan, progression-free survival was longer for biomarker-positive versus biomarker-negative samples, a median of 7.2 versus 4.0 months, hazard ratio 0.40. In patients treated with docetaxel there was no such split, 4.1 versus 4.4 months, hazard ratio 0.94.
AlexSo it's specific to the drug, not just prognosis.
MayaRight, and the interaction between treatment and biomarker status had a P value of 0.0098. Here's why it's an AI result: the paper says conventional pathologist visual scoring of TROP2 by H-score was not associated with centrally confirmed objective response, progression-free survival or tumour size reduction in the TROPION-PanTumor01 cohort, where cases were split at a median H-score of 260 or above versus below.
AlexSo on that trial, with that scoring method, the visual read showed nothing.
MayaThat's what the paper reports. The limits are real. It's a retrospective analysis of samples from existing trials, not a prospective trial of biomarker-guided treatment, and the paper says the marker may be predictive.
Health, science & medicine — Study of 28,870 AI mental-health sessions finds 11.7% ended in escalating distress
AlexA much less comfortable health number, in JMIR Formative Research. Researchers looked at 28,870 sessions from 4,463 users of an AI mental health app and sorted them into five trajectories.
MayaWhat were they?
AlexStable at 26.6%, reduced distress at 24.4%, transitional processing at 21.5%, persistent distress at 15.8%, and escalating distress at 11.7%.
MayaIs there a more hopeful figure?
AlexThere is. Among sessions that began in distress-associated states and had duration data, 73.3% showed reduced-distress trajectories by session end.
MayaHow much weight can we put on the measure?
AlexNot much. The distress score is the product's own classifier, not a validated self-report instrument, and all five authors are affiliated with the company that makes the app, so these are its own numbers. They say future research with self-report validation is needed.
Transition
MayaA short stop in policy before the money.
Policy, regulation & law — House Democrats' first-100-hours plan would end Medicare's WISeR pilot of AI-supported coverage review
MayaNextgov reports that a House Healthcare Working Group memo identified ending the Medicare WISeR model, which tests AI-supported coverage review, as a policy Democrats should advance within the first 100 hours of the next Congress.
AlexIs there a bill behind it?
MayaAn existing one. The memo names the Seniors Deserve SMARTER Care Act, which would ban the pilot. Democrats are set to control both chambers.
AlexWhat's actually new?
MayaThe timetable. The memo came out in September, and we covered a WISeR vendor's warning to the agency in an earlier edition, so this is an update. No bill has moved, the next Congress hasn't convened, and Nextgov is the only source.
Transition
AlexAnd now the compute money, which had a loud day.
Compute, chips & infrastructure — TSMC board approves up to US$500 million each for Anthropic IPO shares and OpenAI preferred stock, plus US$40.9 billion of capital appropriations
AlexThis one is in an SEC filing, which is why it's worth saying out loud. TSMC's board approved an investment of not more than $500 million to buy Anthropic common shares at the IPO price, and not more than $500 million to buy OpenAI preferred stock.
MayaThe foundry is buying into both of its biggest AI customers.
AlexIn the same filing. It also approves capital appropriations of approximately $40.9 billion for capacity, packaging, fab construction and next year's research and sustaining spend.
MayaIs this the circularity people have argued about all year?
AlexIt's that argument, now in a filing. TSMC sells the capacity and holds stock in the buyers of it. CNBC reports Anthropic agreed to sell $26 billion of a separate share class at the IPO price to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC.
MayaWhat isn't done yet?
AlexThe purchases. These are board approvals and caps, not completed investments, and the Anthropic piece depends on the offering actually pricing. The company says all of this in its own filing, so it isn't independently verified.
Transition
MayaWhich brings us to the offering everyone is watching, and a lawsuit about coding agents.
Deployment & impact — Anthropic launches IPO roadshow at $70.00 to $78.00 a share for 825,000,000 Class A shares, with pricing planned for 18 November
MayaThe range is filed. Anthropic says the offering is 825,000,000 Class A shares, with the price expected to be between $70.00 and $78.00 a share. It has applied to list on Nasdaq under the symbol ANT.
AlexWhen does it trade?
MayaCNBC reports that, according to people familiar, Anthropic plans to price on November 18th and trade the day after, with retail allocation targeted at about 20%, against a typical 5% to 10%. At the high end the company would net up to $64.35 billion and could be worth up to $1.405 trillion fully diluted, against the $965 billion valuation announced in May.
AlexWhat do the financials look like underneath that?
MayaLarge and lopsided. CNBC cites run-rate revenue of roughly $81 billion at the end of October, up $10 billion from the previous month, and $33.24 billion of revenue for the nine months to the end of September, against a $104 billion net loss over those nine months, most of it from change in fair value of convertible notes.
AlexAnything on the founders?
MayaTechCrunch reports each of the seven co-founders will sell 3,378,378 shares, and that the founders will control 51.1% of the votes. This is an update on yesterday's edition. The range isn't final, it's a filed range before a roadshow, and the figures come from the company's own filing, so they are not independently verified.
Deployment & impact — Point72's Cubist alleges a former analyst ran 4,117 Claude Code and Codex queries to distil colleagues' research and code
AlexAnd finally, a verified complaint about coding agents and insider theft. Cubist, which describes itself as the quantitative-trading arm of Point72, alleges a former analyst issued 4,117 queries across 2,416 sessions with Claude Code and OpenAI Codex.
MayaWhat does it say he did with them?
AlexIt says he directed Claude Code to read and extract confidential information from 14 different codebases, representing nearly every repository of his team, and that he issued over 600 prompts using exhaustive-read language to distill the team's code into portable markdown documents.
MayaWhen was he running them?
AlexThe complaint says nearly half the queries took place outside normal working hours or at the weekend, and alleges the last step was to extract the files and delete the AI usage records.
MayaHow much of this is established?
AlexNone of it. These are untested allegations, not findings. We read the complaint itself, we have not seen a response from the defendant, no court has ruled, and it is our only source.
Outro
MayaThat's The AI Edge for today. A lot of what we read out came from labs and vendors describing their own work, so weigh it as that.
AlexThe full edition is on the site, with every source listed beside the claim it supports, so you can check the figures yourself rather than take ours.
MayaOur voices are AI-generated.
AlexListen in tomorrow for the next edition.