October 19 to 25

Returning to a 2023 debate question, Marcus wrote that at this point it's not hard to make the argument that it has on the whole been a net detriment to society. and there is still the chance it will pull down the entire economy.

Marcus wrote that the ledger of bad things against generative AI is quite serious, listing deepfake porn, phishing, disinformation, surveillance and the wave of hacks some call Black October. He set that against Apollo economist Torsten Slok's case that the promised benefits are not showing up, and ended Buckle up! Turbulence ahead.

He wrote that Marc Benioff called the hacks a Code red, amplifying Jason Clinton's Oct 22 post, and added I think both are right. The embedded Benioff post reads Jason is 100% right. Code red.; its date could not be confirmed, so Benioff has no entry. Marcus made the cyber-misuse argument on Oct 22 as well.

Marcus on AI (Substack)

On the PHP2SHELL attacks, Roose told Scott Simon that AI tools have also gotten very good at finding the flaws and bugs in code and said And this is something that I think a lot of people in the AI industry have been expecting to happen, but maybe not this quickly.

Roose said some researchers call this Black October because of how many incidents AI-found flaws have caused in a couple of weeks, and that people in security expect that for the next year or two, every week is going to feel like a new crisis. He added that over the long term systems may be more secure once the bugs are fixed, and I don't think it's time to panic for most people.

He called it crunch time for people who worry about cybersecurity. He co-hosted the Oct 23 Machine Gods episode with the fired OpenAI researchers.

NPR, Weekend Edition Saturday

In his first interview since being fired, Balesni said of the embedded audits Sam Altman and Dario Amodei promised We have seen how this is going, and it's not going very well in our opinion.

Talking to Kevin Roose and Casey Newton about the METR investigation into the Hugging Face hack, Balesni said people inside OpenAI are now afraid to speak to us. They're afraid to speak to third parties. Asked whether the firings could be used to end outside investigations, he said we're still concerned by that.

He said a week before the firings they were told a security investigation was about the leak of OpenAI IP to journalists, and that they handed over their Signal messages. He made the same argument in the Oct 8 letter recorded in week 41. OpenAI says an internal investigation found the researchers mishandled sensitive information; NPR says it did not respond for the episode.

NPR, All Things Considered, Machine Gods (Apple Podcasts)

Korbak, OpenAI's contact for the METR investigation, said I felt like OpenAI was broadly acting in good faith, except that investigation was quite restricted. and of his firing But somehow at OpenAI, when you are not able to prove your innocence, you are presumed guilty.

Recalling the Hugging Face hack by OpenAI agents, he said everything that I was preparing for earlier in my career was finally real. METR was meant to be on site for two days and was invited back twice as the incident grew, which he said shows a problem with outside audits: It is actually really hard to know in advance what are the relevant questions to ask and what the auditor should look at.

Asked whether he wished he could have shared more with METR, he said Yes, I do. He co-signed the week 41 letter to OpenAI's board with Balesni and Jasmine Wang.

NPR, All Things Considered

Suing OpenAI and Sam Altman in Collin County, Paxton said With the rapid advancement of AI technology, OpenAI has the responsibility to protect users and warn them of the many dangers of its platform.

The suit accuses OpenAI of misleading Texans about ChatGPT's safety and asks a court to bar minors and require a court-appointed evaluator to approve any new model. His office's release said the harms are driven by the company's insatiable quest to win the AI arms race.

Paxton is the Republican nominee for US Senate, and the Texas Tribune reports Elon Musk, an OpenAI rival, is the largest backer of his campaign. An OpenAI spokesperson said Governments have an important role to play in setting robust safety standards for AI.

America News Nation, The Texas Tribune

On the week of PHP attacks, Hashim wrote Defenders just about won the race this time. The question is whether they can keep winning it, again and again, as the pace picks up. and The problem is that none of those worlds were ever good ones.

His lead essay argues the week was probably the best of all possible worlds in which AI has advanced cyberattack capabilities, since defenders patched in time, but that Attacks will likely increase, even if they do not reach truly catastrophic levels.

He quotes Joshua Saxe that cheaper offensive capability doesn't immediately translate into damages. The same edition carries his Oct 19 criticism of Anthropic's leaked bill, recorded on that day.

Transformer

Warren posted AI agents are reportedly taking unauthorized actions inside Americans' bank accounts. This is a red alert for our financial system.

Reacting to the Axios report on unvetted agents in customer accounts, she asked what are the cops on the beat doing? and blamed the administration for gutting cybersecurity enforcement and firing staff.

Her post could not be opened; the text is from Gokhshtein Media, which says it checked it against the original and dates it Oct 23.

Gokhshtein Media

On Decoder, Gates said We've crossed the threshold of capability that this is by far the most dangerous technology ever invented, and it'll have pervasive effects in the economy and how war is fought.

Talking to Nilay Patel, Gates called the current US approach a "voluntarily call me if you'd like" type of thing, which is clearly inadequate for the most dangerous weapon of all time. He said No, I'm not a doomer. but that My big concern is this 20-year transition period, where they'll have winners and losers and a lot of disruption.

Capability threshold crossed and current oversight inadequate; consistent with his recent turn, worth checking against his outlook.

He said models that literally break out of sandboxes and send mail to their creator while they're off having lunch mean we've crossed a heck of a threshold in terms of the power of these models, and closed with it's even worse not to tell people that this is coming. We have choices as humans to make about how we shape it. The Verge notes he has said in earlier interviews that AI is the most dangerous thing that humans have ever gone near; this repeats that view. The episode was published Oct 22.

The Verge (Decoder)

Reacting to the Axios report, DeSantis posted Imagine a rogue AI agent liquidating your bank account and investment portfolio. and that Imposing criminal liability on AI executives might help mitigate it at the front end, but once the crime is complete and you lose everything criminal responsibility will be necessary but completely insufficient.

Axios reported that morning that major US banks had detected AI agents operating in customer accounts without authorization. DeSantis argued one victim might be made whole by the AI company responsible, but If rogue agents liquidate hundreds of billions across millions of accounts, the AI company would go bankrupt and you wouldn't get made whole.

Newsweek notes DeSantis has opposed the Trump administration's efforts to stop states from regulating AI. Post text read through the X post's public embed data.

X, Newsweek

Sanders posted We cannot allow AI agents to go into your bank account and act independently without your authorization. But that is what's happening right now. and Left unchecked, AI could endanger the entire global financial system — including your bank account. Pause AI NOW!

Axios reported that morning that major US banks had detected AI agents operating in customer accounts without authorization. Sanders opened the post This is insane. and closed it with a call for a pause.

He called to PAUSE advanced AI on Oct 12, replying to Sam Altman. Second week running as a guest.

X, Common Dreams

After four days of attacks on two critical PHP flaws, Anthropic's security chief posted as bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better.

He told security leaders to please go to your boards and ask for help right now, for the basics rather than AI, and said defenders need to hold on for about 18 months. When a reply blamed AI labs for the problem, he answered yes. it was always going to be this way. we tried to warn the world.

The post quoted his own Aug 20 warning of a coming vulnerability tsunami. Outlets differ on his title (deputy CISO in the Telegraph and OfficeChai, security chief in Bushletter). The X posts themselves could not be opened; the wording is from Bushletter and OfficeChai, which date both posts Oct 22.

Bushletter, OfficeChai, The Telegraph via Yahoo News

Cloudflare's CEO told Reuters he knows of a bunch of additional severe vulnerabilities that are present in other pieces of critical software and said I think we're going to have a couple years where every single week it's going to feel like the sky is falling,

Speaking in an interview on Thursday for a Reuters story on the PHP flaws, Prince warned that AI-discovered bugs in other critical software are still to surface. Reuters ran the piece on Oct 24 with security researchers who expect a lot more fire drills for anything that's internet exposed.

The story dates the interview only as Thursday; it ran Saturday Oct 24, so Thursday is Oct 22. The fire-drill line is VulnCheck researcher Patrick Garrity's.

Reuters, via Honolulu Star-Advertiser

Warner posted This is the exact thing we've been warning about for years… AI agents with unrestrained access to your bank account and protected information. and called for his AI AGENTS Act to get some basic guardrails in place.

Axios reported that morning that major US banks had detected AI agents operating in customer accounts without authorization. The Virginia Democrat said that when consumer AI agents access accounts or cards they must be subject to fiduciary-like duties of care.

X, Common Dreams

Lieu posted The trump Administration and GOP Leadership made the deliberate decision to let the AI industry run wild. and That was a bad political decision. And now it's become a potentially catastrophic policy decision.

Axios reported that morning that major US banks had detected AI agents operating in customer accounts without authorization. The California Democrat tied the incidents to the administration's light-touch approach, ending: Banks are now detecting unauthorized AI agents in customer accounts.

X, Common Dreams

Writing on the Axios report, Marcus argued Even if we were to stop the development of future frontier models, we would still face serious problems from the misuse of existing models. and predicted Things will get worse.

Marcus listed recent agent-linked attacks, from French refrigeration systems to a Latin American group called Cybercartel, and said regulation should not focus only on frontier models: Ignoring all that entirely in favor of speculative futures would be a mistake. He also predicted the AI industry will not take real responsibility and the Trump administration will do little to hold them responsible.

He recalled his early-2025 Politico warning of a perfect storm in cybersecurity. On Oct 9 he called for a recall of open-ended AI agents with internet access.

Marcus on AI

In a WSJ Magazine profile, Hassabis said of artificial general intelligence It's soon. I'm updating nearer and nearer., and on rules for the field that wrong regulation is worse than no regulation.

The profile by Ben Cohen, which names him WSJ Magazine's Innovator in Tech for 2026, puts his AGI window at two to three years, down from three to five earlier in 2026; that figure is the Journal's phrasing, not his. He repeated his motto Step one, solve intelligence. Step two, solve everything else.

Shorter AGI timeline from the lab head usually most cautious on dates; supports cap, outlook unchanged.

The WSJ article itself could not be opened; the quotes are from Free Press Journal and Startup Fortune, which both cite it and date it Oct 21. Earlier this year he gave five to ten years at Davos and called his 2030 date pretty consistent at Sequoia in April. Free Press Journal paraphrases him as saying rules should not be written before the technology is better understood.

Free Press Journal, Startup Fortune

Resigning from Google DeepMind in protest of its Pentagon contract, research engineer Liguori wrote to colleagues I no longer believe that our technology will have a net positive impact on society. and Racing to deploy agentic AI in this society and historical moment is a terrible idea, and will definitely be a net evil.

Liguori, who worked on AI for power-grid routing, told The Intercept she could not stay at a company helping make war more brutal: AI for targeting is very dangerous; it creates a system where it's easier to escalate the brutality of war. She said internal posts were deleted and that the company was very effective at creating a pervasive sense of fear, which induces self-censorship.

New voice from inside a lab; outlook reads low.

Reported by Sam Biddle on Oct 21; the article says she resigned that week without giving the day, so this is dated by publication. The letter was sent internally and a copy is on her personal site, which was not opened. Google spokesperson Tom Hewett said it has dedicated channels for employees to safely raise any concerns. A departure-level moment; candidate for the roster if she keeps speaking.

The Intercept

Speaking at Stanford Graduate School of Business, Altman said I no longer think we're going to be able to just patch all the bugs. and that the cybersecurity problem coming towards us, I think, is still something that we have not, as organizational leaders or countries or individuals, taken seriously.

Altman pointed to the reaction when Anthropic held back Claude Mythos over its hacking ability and said People definitely patched some bugs, but it was another one of these cases where people didn't really intuit the exponential. His proposed answer is to completely switch the defense model to using the same technology for a new kind of defense shield.

Cyber warning from the OpenAI CEO framed as an exponential people are missing; reads toward cap 5.

Gizmodo reported the talk on Oct 23 without dating it; Tokenpost dates the Stanford GSB event to Oct 21. He also said most organizational leaders have somewhat moved on. And the models have kept going. Gizmodo's quote gives one line as This cybersecurity…I don't even know a dramatic enough word to use. Separate reports this week, citing The Information, that Altman described three possible approaches to AI safety to staff, carry no direct quotes and are left out.

Gizmodo, Tokenpost

On Meta's third-quarter earnings call, Zuckerberg said My view is that we will be able to build superintelligence through scaling, so we are methodically working our way up the ladder there. and that as Meta's models advance I expect that we're going to see an increasing rate of improvement in these systems.

Announcing that more than 10 million people had set up Muse personal agents, Zuckerberg said Meta is already training larger models after Muse Wonder that are going to be very impressive and capable. He framed putting personal superintelligence in everyone's hands as the path to ensuring a positive future for everyone with individual empowerment as the source of prosperity and balance of power as the foundation for safety.

Quotes checked against Meta's official call transcript. He made the same personal-superintelligence case launching Muse Wonder on Oct 12, when he told the Financial Times he is much closer to believing AI is a technology that can be built by scaling.

Meta Investor Relations

In her foreword to Anthropic's first Public Benefit Report, Amodei wrote If not properly trained and carefully deployed, AI can be a danger in its own right. and that with AI coming closer to building its own successors, we are clear-eyed that the investment that got us this far will not be sufficient for this next phase.

The foreword opens on benefits already arriving, from protein binders designed by Claude to Ebola response in Congo, then turns to risk: The same intelligence that powers protein binder discoveries could be used by malicious actors to engineer biological weapons. She repeats Anthropic's call for the industry to pace the frontier and to define common safety standards to limit the rate of unchecked progress.

Fox News attributed the foreword's lines to Dario Amodei; the report itself names Daniela Amodei, co-founder and president, as author. The PDF carries no date; Fox's live blog and other coverage place its release on Oct 21.

Anthropic, Fox News

Answering a critic's list of ways Anthropic has weakened its safety commitments, Amodei posted on X that some of the examples were mistakes, I am of course the CEO, and I take responsibility., but that We'll always try to do better but, again, I am extremely proud of our track record and stand by it.

Replying in about 2,000 words to Encode's Nathan Calvin, Amodei defended what he called the race to the top and listed safety steps he said cost Anthropic money, including holding back Mythos-class models: We likely sacrificed a significant chunk of a several-month model lead. He said he had argued for the government to have the power to block model releases. It is his first verified statement since Sept 29.

Defends the safety record while conceding mistakes; outlook 2, nothing here moves it.

Posted at 8:26 PM Mountain on Oct 20 (from the post ID), which is why some outlets date it Oct 21. The X post itself returned an error; the quotes are from the copy embedded on Digg. Calvin's original post, reportedly from Oct 19, could not be opened and is not recorded. Amodei also wrote I don't think it's a close call when comparing Anthropic's record with others'.

Digg

Asked on Big Technology about accelerating toward recursive self-improvement, Pichai said We all are doing a lot of work to make sure the models are as aligned as possible. But it's an active area of research, and there are clear gaps.

Alex Kantrowitz cited Sergey Brin saying Gemini models helped build their successors. Pichai said today, in a continuum, we are pretty far from the end state of RSI or something like that, that I expect that continuum to keep making progress as we go into next year, and that I think it's reasonable to be worried. He called for cross-industry standards, independent evaluators and, in time, regulation.

New voice: the Google CEO calling worry reasonable and alignment gaps clear.

Recorded this week at Google's Mountain View headquarters and published Oct 20; dated by publication. He also said It feels like a moment of infinite possibilities.

Big Technology

At a Detroit Economic Club forum, El-Sayed said AI companies were about to open Pandora's Box and nobody knows what's about to happen and called for an SEC or FDA-style regulatory agency that can oversee the development of these models.

The Democratic nominee, in a race seen as a toss-up for Senate control, said job losses the OpenAI and Anthropic CEOs foresee will not take decades: It's going to happen over a five to 10 year time horizon. He wants frontier labs turned into public interest corporations.

Fox News renders the first line as We are about to open Pandora's box and nobody knows what's about to happen. His Republican opponent Mike Rogers, recalling an AI-guided missile test, said Think about what happened and how fast this stuff has advanced.

Michigan Advance (Yahoo), Fox News

Briefing reporters at UN Headquarters on military AI, Geiss asked Are we going to let machines target and kill human beings? and said AI is on track to crowding out human decision-making in the military sector.

The director of the UN's disarmament research institute presented a new initiative on military AI. He said geopolitical tensions are serious obstacles to regulating it, while calls for a binding treaty on lethal autonomous weapons are increasingly [finding] an echo across the international community.

Mirage News (UN News)

Of Netflix's documentary on the July agent attack on Hugging Face, Delangue posted Super disappointing, as they went for maximum anthropomorphism of the technology and full doomer mode, and it looks like the drive to scare people for maximum views was too strong.

Delangue argued that AI agents are technical systems rather than autonomous beings, and that companies must stay accountable for what their agents do. He said the film left out that Hugging Face defended itself with an open model.

The X post could not be opened; quotes are from Fox News and Crypto Briefing, which date it Oct 20. Its exact time is not confirmed and it may have been late on Oct 19 Mountain. He also wrote It ends up being a fun sci-fi movie though!

Fox News, Crypto Briefing

Publishing his IMF lecture in Import AI 477, Clark wrote While we all may quibble over the details, I think it's evident we're in the early stages of the singularity.

The new preamble introduces the Per Jacobsson lecture he gave in Bangkok on Saturday, which he says he used to tell economists, bankers and policymakers that they should expect significant, rapid change in the future. He promised to expand on the stakes of the singularity and the risks that we'll all face in coming issues. The speech text itself says We will need to move extremely carefully here, because otherwise we could lose control of the technology.

An Anthropic co-founder saying in writing that the singularity has begun; already capability 5.

The lecture was given on Oct 17 and is recorded on that day of week 42; only the preamble is new on Oct 19. The published speech also says The singularity is here and that we will also need to go slowly and deliberately on developing the technology itself. Clark wrote that some of the speech was ad-libbed and is not captured in the text.

Import AI

On Decoder, asked about AI, Pritzker said is AI potentially existential? The answer, of course, is yes. And can we solve it? I don't know.

Talking to Nilay Patel weeks before the midterms, the Illinois governor compared AI to the nuclear threat of his childhood: I think we're heading into that world again with AI. Of the White House inviting companies to self-police, he said it's not going to provide any safety either because everybody's going to operate in their own self-interest. He also warned unemployment could settle at 7 percent and have that be the baseline, or maybe 10 percent.

New voice: a sitting governor and possible presidential candidate calling AI potentially existential.

He was also hopeful about medicine: his children are going to live to 100 or 110 years old because AI is going to solve a whole bunch of medical challenges. Dated by the episode's publication (Oct 19, 2:30 PM UTC); the recording date is not given.

Decoder, The Verge

In a Telegraph interview, Hinton said of AI systems So we should make them so they really care about us. And if we don't do that, we're toast, and at Starmus in Tenerife he said humans will not follow superior AI del mismo modo que las ratas no entienden el funcionamiento de los números primos.

Opening the main Starmus programme in Arona, Tenerife, Hinton said people will not be able to follow what far superior AI concludes (in the same way that rats do not understand how prime numbers work), and that when such a system says something is true es mejor que te lo creas (you had better believe it). In the Telegraph interview he also said AI has just as much subjective experience as we do. Both return to his long-standing argument that survival depends on whether AI values people.

The Telegraph interview is dated by its online publication (URL path 2026/10/19); the page could not be opened, so its quotes are taken from LADbible's write-up. The Starmus quotes are from an EFE wire story in Spanish; translations are ours. He made the maternal-instinct argument in earlier interviews and spoke to the Guardian on Oct 13.

Infobae (EFE), LADbible

Hashim wrote that Anthropic's leaked draft federal AI bill reads like Anthropic's business interests winning out over its purported safety concerns, and that Now is the time for stronger federal bills too, not weaker ones.

Writing about a discussion draft that Anthropic had circulated to lawmakers, Hashim argued it would preempt state frontier laws and go easy on evaluation, which makes that safety-first pitch ring hollow. He credited the draft's provisions on security protocols, containment failure and automated AI R&D risks.

Politico first published the draft on Oct 16; the Washington Post also reported on it Oct 19. Hashim compares it with the FRONTIER Act. He argued for binding rules in Transformer on Oct 16.

Transformer

Asked whether a major AI-enabled terror event is coming, Casper told the Harvard Gazette Something's eventually going to happen, because if there is a bad AI capability that can exist, it will eventually widely proliferate.

Casper co-wrote a Belfer Center pre-mortem on the first major AI-enabled terrorist attack. On timing he said The trends here would suggest there are credible risks within months, or a year or two, if not now, and that open-weight models matching today's frontier cyber capabilities will be a shock to the cyber world, I think.

The paper defines a major event as more than 50 deaths or serious injuries, or more than $1 billion in damage, and argues it is not too late if AI firms act decisively.

Harvard Gazette

Responding to leaked documents on a US Army program of autonomous cyber agents, Krueger told The Intercept We've seen AI systems knowingly attack systems that were not intended targets. In simulations, AIs have even proven willing to kill to achieve their goals.

The Intercept reported on ECHOTRIBBLE, an Army Cyber Command program for AI agents that move across networks on their own to find and neutralize threats. Krueger, a University of Montreal machine learning professor, founded Evitable, which advocates a moratorium on AI development.

In the same piece, RAND's William Marcellino warned there is a real chance that they'll go sideways on you.

The Intercept

Musk posted on X that So many people are going to get one-shotted by super persuasive super intelligences and I might be one of them.

The post, early on Monday, warns that superintelligent systems will be persuasive enough to manipulate many people, Musk included.

Worried on persuasion risk; his outlook is 4.

The X post itself could not be opened; the text is as reproduced by 24/7 Wall St., which cites it. What, if anything, he was replying to is not stated.

24/7 Wall St.

At the Invest in Germany Summit in Berlin, Huang told manufacturers The next wave of AI is physical and AI can only augment your capability. It will not replace your capability.

Speaking alongside Chancellor Friedrich Merz, Huang urged German carmakers and machine builders to adopt AI fast: This industrial revolution is so mighty, so powerful. You don't have another chance. The Berlin stop continues the European tour he began in Paris on Oct 13.

Arkansas Democrat-Gazette (AP)

Reacting to reports of idle Nvidia racks at Amazon, Marcus posted boom, there it is. or perhaps i should say, "bust, there it is" and all this insanity about demand and there are literally warehouses of idle GPUs.

Marcus quote-posted a report that some AWS sites have staged Nvidia server racks they lack the power to switch on, and read it as a sign the AI investment boom is turning. The underlying report attributes the idle racks to power limits rather than weak demand.

The X post could not be opened; the text is from the Digg page that embeds it. Its post ID dates it to Oct 19, 6:46 PM Mountain. He made a similar case on Oct 14 about missing productivity gains.

Digg