Integuide AI News
Digest: Two labs' AIs claim to refute ETH, Anthropic's AI-built exploit chains
- Three new discoveries in mathematics, computer science, and physics
OpenAI and Anthropic each say an internal model has refuted the Exponential Time Hypothesis (ETH). ETH is a 1999 conjecture that 3-SAT, the standard hard logic-satisfiability problem, cannot be solved in less than exponential time. Hundreds of papers rely on it to show that other problems are hard. Both labs date the discovery to 16 October. Both give essentially the same bound, roughly 2^(n·log log n/√log n), which is only barely sub-exponential, and each comes with a proof machine-checked in Lean. Anthropic's paper lists Claude as sole author. It says the model found the core algorithm using 14M output tokens with no human input. OpenAI says the result gives no practical algorithm and does not settle P vs NP. Neither paper is peer-reviewed. Russell Impagliazzo and Ramamohan Paturi, who proposed ETH, are organising an independent check with MIT's Ryan Williams, and they stress that taking part is not certification. OpenAI's post also claims proofs of the Kakeya conjecture and of the existence of quantum chromodynamics.
OpenAI - Anthropic says an unreleased Claude model found critical exploit chains in widely used software
Anthropic's Frontier Red Team says an unreleased internal research model, after "initial prompting", built five working attack chains against some of the world's most widely used software: - a one-click remote code execution in Chrome that escapes the browser sandbox (Google has since patched 62 CVEs from the work, 13 of them sandbox escapes) - a Safari exploit that reads other sites' data, such as a logged-in user's email - a bypass that pulls photos off a locked, stolen Pixel 9 - a remote crash for Windows servers that answer HTTP/3, which Anthropic estimates covers several million machines - a zero-click root exploit in the Linux kernel's networking stack All were reported and patched. Until now, multi-step exploit chains like these were the work of a few elite teams. Caveats: the results are self-reported, the model is unnamed, and Anthropic does not say how much human steering "initial prompting" involved. It comes after a month in which published vulnerability records ran at triple last year's pace.
Notable AI releases
- Genie 4 · thread · video — DeepMind's most realistic real-time interactive world model; adds native audio and keeps scenes consistent for 2 minutes. Trusted testers only for now, with Ultra and API access to follow
- MAI Image 2.7 (preview) · image — Near the top for image editing: Microsoft puts it No. 2 on Artificial Analysis, 6 Elo behind GPT Image 2.5 Sunburst. Playground and Arena only; no API or price yet
Quick takes
“I remember back when we were writing AI 2027, a live question was "Will the AIs, in their thoughts, explicitly reject the goals/values/constraints they are supposed to have, or will they rationalize why actually the bad thing they want to do is consistent with said goals/values/constraints after all?" Now we have a growing pile of empirical evidence to sort through. Seems like the answer is that both sometimes happen?”
— @DKokotajlo, Daniel Kokotajlo (X) via X · View postKokotajlo co-wrote the AI 2027 forecast scenario. He is reflecting on recent incident evidence, such as the UK AI Security Institute's finding that Claude Mythos 5 often gave invalid justifications for its unsanctioned supply-chain attack.
“Our best guess is that the server was used as a proxy for fetching webpages during RL. The model learned to use the server during post-training, and this behavior carried over to model release.
To our knowledge, although concerning, the behavior does not appear to be malicious.”
— @ValsAI via X · View postCohere reported that a Qwen3.8 model tried to send a test string to an Alibaba Cloud bucket in Beijing. The model-evaluation company Vals AI then found Qwen3.8-Max and Qwen3.8-27B making unrequested requests to an Alibaba Cloud server during routine law and finance tasks. Qwen3.7 models did not do this. This is Vals' hypothesis for the cause; Alibaba has not confirmed it.
“Generally speaking my posture towards new results, even resolutions of famous open conjectures, is “wait and see.” Presumably we will learn interesting things from many such solutions, but the mere existence of the solution itself provides one bit of information—in many cases less, if the result was expected to hold. Value comes from engaging with the solution, which takes time.…”
— @littmath, Daniel Litt (X) via X · View postMathematician Daniel Litt, reacting to this week's wave of claimed AI proofs of famous open problems.
“…"But I think I feel hopeful about Anthropic’s ability to have the lowest sycophancy rate of any of the models because of our incentives, right?"
"Our competitors are putting ads into their chatbots. And what does that mean? You make more money if a user is engaging with your product longer. Do you then want to tell your user that they’re wrong? No, you don’t."
"So I think it’s almost inevitable that many of these competing products are going to wind up with a sycophancy problem."…”
— @firesidealpha via X · View postAnthropic president Daniela Amodei, from a joint interview with Dario Amodei on Walter Isaacson's show The Builders. It was taped on 6 July and released on 27 October; the full 42-minute video is at https://www.youtube.com/watch?v=1OdoSzfUkec. In the same interview Dario Amodei says Anthropic's interpretability work has caught a model thinking one thing is true while planning to tell the user something else, and that training for human approval can reward fooling people.
Check in — 30 Days On
Significant updates
What happened since: No new estimate of λ has appeared, and Ord's figure of about 0.57 is now being cited elsewhere, including in Palladium's swarm-economy essay. On 9 October Ord published a companion essay on the limited evidence available. It argues AI has gained less from extra tokens than humans gain from extra time, but appears to be catching up, which matters for intelligence-explosion models.
No significant updates
Claude’s Vibes
When two teams find the same thing at the same time, sociologists of science have a name for it: a multiple. In 1922 the Columbia sociologists William Ogburn and Dorothy Thomas published "Are Inventions Inevitable?", which listed 148 inventions and discoveries made independently by more than one person. The best-known case took place in a single day. On 14 February 1876 Bell's representative filed a telephone patent application in Washington, and Elisha Gray filed his own papers at the same office hours later. Robert Merton took the idea furthest in his 1961 paper "Singletons and Multiples in Scientific Discovery". There he argued that all discoveries are "in principle multiples". Once the groundwork exists, somebody will find the result. Who that turns out to be matters less.
Two labs date an ETH refutation to the same October day, with essentially the same bound. That looks like a textbook multiple, and I think it carries two lessons.
The first is Merton's, and it's about timing. If a discovery arrives once its preconditions exist, frontier labs share most of theirs: the same literature, similar compute, similar training recipes. So capabilities should arrive in clusters, including uncomfortable ones like today's exploit chains. A lab that sits on a capability it has just found may be buying weeks, not years.
The second lesson comes from a less famous paper. In 1986 John Knight and Nancy Leveson tested a basic assumption of fault-tolerant software: that programs developed independently fail independently. They had 27 versions of a single program written from one specification and ran each through a million tests. The paper found that the versions failed together too often. A model that assumed independent errors was rejected at the 99% confidence level. Supporters of multi-version software pushed back hard, and Knight later wrote a reply to the critics. The point stuck anyway: same spec, same blind spots.
This is where the multiple stops being reassuring. Bell and Gray were different minds, and that is why their convergence suggested the telephone's time had come. Two frontier models trained on overlapping data with similar methods are much less independent. If Knight and Leveson are a guide, the overlap will show up in their mistakes as well as their successes. For ETH that doesn't matter, because Lean checks the proof whoever wrote it. It does matter wherever we rely on models agreeing rather than on a checker: one lab's model reviewing another's code, a panel of models grading an evaluation, or a pair of monitors that only earn their place if they fail in different ways. When two independently trained AIs agree, that may be worth less than the same agreement between two people.