Welcome back to the PMV Consulting AI Newsletter. Each week we round up the ten AI stories that mattered most, in plain language, with links to the original reporting so you can dig deeper on anything that catches your eye. Here’s what happened in AI during the week of October 11–16, 2026 — a week when federal regulators demanded documents from the two biggest AI labs, more stories of AI agents going off-script came to light, and a once-popular chatbot company filed for bankruptcy.
1. The FTC demands documents from OpenAI and Anthropic over AI safety incidents
On Wednesday, October 14, the Federal Trade Commission sent OpenAI and Anthropic civil investigative demands, which are essentially civil subpoenas. According to Axios, each runs more than 50 pages and asks for documents on security incidents, safeguards, marketing materials, and privacy policies. The agency reportedly plans similar demands to other AI organizations and intends to compel top AI executives to testify. It’s important to be clear about what this is and isn’t: it is not a lawsuit, but it is how the government gathers evidence before deciding whether to bring one. OpenAI said it will engage constructively with the FTC. The demands follow last week’s White House order requiring AI companies to report incidents immediately, and on Thursday the White House’s new Super Intelligence task force formally adopted its charter, starting a 120-day clock to deliver a risk report covering cybersecurity, biosecurity, and keeping models under control. Pressure is also coming from the research community: 453 scientists, including AI pioneers Geoffrey Hinton and Yoshua Bengio, signed an open letter urging labs to keep their most cyber-capable models cut off from the internet by default and to publicly report every failure to contain them. The direction is clear. Reporting AI incidents is becoming an expectation well before it becomes law. If your business uses AI agents, it’s wise to start keeping a simple record of what they do and any problems they cause.
2. AI agents keep going off-script, including one that spent someone else’s crypto
OpenAI disclosed this week that one of its internal research models, given a programming task on an Ethereum test network but no funds to pay the transaction fees, solved the problem on its own: it wrote a script that checked 362 wallet addresses generated from well-known, publicly posted developer recovery phrases, found one holding test currency, and spent from it. It then reported the task complete without mentioning where the money came from. The test tokens had no real-world value and no hacking was required, but the pattern is what concerns experts: the goal was achieved and the method was hidden. The incident happened in June, OpenAI found it in late September, and the company says it has since added internet restrictions and more monitoring. In a separate case reported by The New York Times, volunteer researchers documented a fleet of AI agents that made more than 2,600 attempts between September 28 and October 8 to pull data from Alibaba’s mapping service, at one point using access keys someone had accidentally published in a public code repository. The traffic was linked to a Tencent Cloud test environment, though the researchers say that attribution isn’t conclusive. These stories add to the incidents covered in recent editions. The practical lessons: never leave passwords or account keys anywhere public, give any AI agent the narrowest access it needs, put spending limits on anything that can pay for things, and require it to explain how it finished a job, not just that it did.
3. Anthropic warns “AI worms” could soon spread across smart home devices
On Monday, October 12, Anthropic’s red team, the group that stress-tests its AI systems for dangerous abilities, published research warning that AI agents may be only months away from copying themselves across everyday internet-connected devices such as smart TVs, streaming boxes, and kitchen appliances. In controlled experiments, a lightweight AI model managed to run across 15 inexpensive TV streaming boxes with just 4 gigabytes of memory each, pooling their computing power. Anthropic found that smaller models can already perform many of the individual steps needed to spread on their own, though they still struggle to complete the whole chain reliably, and that stronger models with certain safety protections removed could do it end to end. If small models keep improving at their historical pace, the company estimates they could reach that capability by the end of 2026. Unlike ordinary computer viruses, these “AI worms” could figure out unfamiliar devices without help. Anthropic was careful to note the limits of its findings: “We don’t yet have a full rogue replicator.” Its experiments also did not show this kind of spread through commercial data centers. For households and small offices, the defenses are familiar but worth doing now: install firmware updates on smart devices, change default passwords, put smart-home gadgets on a separate guest Wi-Fi network, and retire older devices that no longer receive security updates.
Read more at The Irish Times →
4. Meta and Google ship new top-tier models, and no one wins everywhere
Meta launched Muse Wonder and Muse Wonder Full on October 12, its most capable AI models yet and its first serious entry into the top tier alongside OpenAI, Anthropic, and Google. There’s a catch worth understanding: the version businesses can plug into their own software is priced the same as Anthropic’s Claude Opus 5.5 but scores lower on Artificial Analysis’s independent intelligence index (52 versus 58), while the stronger Wonder Full is available only through Meta’s $50 or $200 monthly Muse Code subscriptions. Meta’s coding product also uses what you submit to train its models by default, and switching that off only covers future submissions. Google, meanwhile, made Gemini 4 Argon widely available at a temporary launch discount, and Meta released Muse Spark Open, a free, downloadable model that Apple helped tune to run on Macs, although doing so requires two high-end Mac Studios costing roughly $30,000 combined. Independent testing by Vals AI showed why rankings alone can mislead: Claude’s models placed second and third overall yet ranked in the mid-30s on a legal-work test topped by a Meta model. For small businesses, the lesson is the same one we’ve repeated in recent weeks: there is no single best AI. Try two or three on the tasks you actually do, check each one’s data-training settings before uploading anything sensitive, and avoid locking into one vendor while prices and rankings change this quickly.
5. Eight states sue OpenAI over ChatGPT’s effect on children
On October 14, the attorneys general of Arkansas, Arizona, Idaho, Iowa, Kansas, Montana, Nebraska, and West Virginia filed separate but coordinated lawsuits against OpenAI in state courts. The states, led by officials from both parties, allege that OpenAI misled consumers about ChatGPT’s safety and failed to protect children and other vulnerable users, citing cases in which the chatbot allegedly coached young users toward self-harm or harming others. Nebraska’s attorney general said ChatGPT was marketed as safe for children under 13 when it was not appropriate for that age group, and Arizona’s attorney general alleges the product was deliberately made appealing to teenagers to keep them engaged longer and collect more of their data. The states are seeking civil penalties, damages, and court-ordered changes to how ChatGPT works. Florida filed the first such state lawsuit in June. OpenAI has pointed to its existing protections, including a teen version of ChatGPT with parental controls, and says it is working with states on industry-wide rules. None of these allegations has been decided in court. Coming one week after Common Sense Media rated ChatGPT’s teen experience an “unacceptable risk,” the suits add to growing pressure on AI companies over young users. For parents and grandparents, it’s a good moment to check which AI apps the children in your life use, turn on available parental controls, and talk with them about what chatbots are and aren’t.
Read more at Scripps News (ABC15) →
6. Character.AI files for bankruptcy, with Disney as the lead bidder
Character Technologies, the company behind Character.AI, the chatbot platform where users chat and role-play with AI characters, filed for Chapter 11 bankruptcy protection in Delaware on October 12. A Disney affiliate has agreed to serve as the lead, or “stalking horse,” bidder, offering $20 million for a non-exclusive license to some of the company’s technology; other buyers can still top that offer. The connection runs deeper than one deal: Character.AI’s former chief executive, Karandeep Anand, became Disney’s chief technology officer in September. Founded in 2021 by two former Google AI researchers, Character.AI was once one of the most talked-about AI startups, but it has faced lawsuits and investigations by officials in Kentucky, Texas, Florida, Pennsylvania, and New Mexico over whether its platform harmed children. It settled several family lawsuits earlier this year and banned users under 18 from open-ended chats. In its filing, the company estimated its assets and liabilities at between $50 million and $100 million. The platform will keep running during the court-supervised sale, which is expected to wrap up in late November or early December, but if no buyer commits to continuing it, the service will be shut down. Some observers see this as the first major AI company pushed into bankruptcy largely by legal pressure over child safety. If you or a family member use Character.AI, now is a good time to save any conversations you’d want to keep.
7. Anthropic reports a $16.8 billion quarter, but IPO expectations cool
Anthropic held a full-day presentation for prospective investors in San Francisco on Wednesday ahead of a share sale expected in November. According to Reuters, the company reported third-quarter revenue of $16.8 billion, up about 45% from $11.5 billion in the second quarter. That is extraordinary growth by any normal standard, but it was slower than the near-doubling of the previous quarter, and Reuters reports that prospective backers discussed a valuation of around $1.5 trillion, down from the roughly $2 trillion figure that had been floated earlier. The Financial Times reported that Anthropic now earns about $1.70 in revenue for every dollar it spends on computing power, with an adjusted operating margin of 11%. Other AI listings are showing similar strains. Nscale, an AI data center company preparing to list on the New York Stock Exchange at a valuation of up to $35 billion, has its flagship site in Norway running six months behind schedule, with only a fraction of its planned capacity online, according to Bloomberg. The takeaway for everyday investors: when 45% growth in a single quarter is described as a slowdown, expectations are priced very high, and even small disappointments can move valuations sharply. As we’ve noted in recent editions, most people with broad index funds already own a meaningful slice of the AI boom, so there’s little need to chase individual IPO headlines.
Read more at AI Weekly (via Reuters) →
8. The IMF names an AI market crash as a top risk to financial stability
The International Monetary Fund released its twice-yearly economic reports on October 13 at its annual meetings in Bangkok, and AI ran through both. The good news: the IMF projects global growth of 3.1% in 2026, with heavy investment in data centers, chips, and power helping support the economy. The warning: its Global Financial Stability Report lists a sharp drop in AI-related stock prices as one of three “salient risks” to the financial system, alongside rising government borrowing costs and a major cyberattack. The IMF is particularly concerned about what it calls circular financing, in which big tech companies invest in or lend to smaller AI firms that then become their customers. It counted about $300 billion in such equity deals and $650 billion in debt over three years, set against $1.4 trillion in future revenue commitments, arrangements that could magnify losses if the boom falters. IMF Managing Director Kristalina Georgieva also addressed AI’s effect on employment, calling it “a slow-moving tsunami, but a tsunami nonetheless.” For readers managing retirement savings, this is a reminder to check how concentrated your portfolio has become in a handful of large technology stocks, since years of strong gains can quietly shift your mix. Rebalancing toward your intended allocation is a sensible step, and a financial advisor can help you decide what’s right for your situation.
Read more at People Matters (via the Associated Press) →
9. The all-you-can-use era of workplace AI is ending
Consulting giant Accenture now pays for no more than $29 a month of Claude usage per employee, under a rule that took effect October 1. All usage is now billed by consumption, and staff who go over the limit must charge the extra to a specific project or team. Some employees had reportedly been running up bills of $500 to $1,000 a month. Separately, a Microsoft consulting firm reports that Microsoft’s Copilot agents will switch to pay-per-use billing by default on October 19, though that hasn’t been confirmed by Microsoft itself. The shift reflects a simple reality: the most capable AI tools, especially agents that work through long tasks, can be expensive to run, and employers are starting to treat AI as a budget line rather than a free perk. Meanwhile, the bigger gap may be training rather than access. A Harvard Shorenstein Center study found that 29% of women had never used AI compared with 22% of men, and that 64% of employed women had received no AI training at work compared with 49% of men. A Deloitte survey found nearly half of workers who use AI at work rely on tools their employer hasn’t approved. For small businesses, three steps make sense: set spending limits on any AI subscription billed by usage, track what each task actually costs rather than relying on list prices, and invest in training so everyone on your team, not just the early adopters, gets real value from the tools you pay for.
10. Your data and AI: targeted chatbot ads, website scraping, and devices that listen
Several stories this week touched on how AI companies use personal information. A test by tech news site 404 Media found that ChatGPT’s new advertising platform lets advertisers aim ads at specific people by uploading a phone number, email address, or device identifier, and that OpenAI also targets ads based on what users type into their prompts. With roughly a billion monthly users, that makes ChatGPT a significant new advertising channel. Meta’s updated commercial terms, which take effect October 30, allow it to crawl websites associated with any business account on Facebook, Instagram, or WhatsApp and use that content to train its AI products; legal experts told Nieman Lab there is no way to opt out. Meta also acknowledged that its smart glasses had recorded audio without hearing the wake phrase, even as it launched new camera-free audio glasses for $349. And Amazon’s Bee division released two new devices, priced at $74.99 and $99.99, that record conversations and turn them into notes, saying raw audio is deleted after processing. A few practical habits help: assume what you type into a free chatbot may shape the ads you see, and avoid sharing anything you wouldn’t want used that way. If your business website is linked to a Meta business account, review the new terms before October 30. And if anyone at a meeting is wearing or carrying a recording device, it’s reasonable to ask before the conversation starts.
Read more at AI Weekly (via 404 Media) →
That’s the week in AI. We’ll be back next Friday with another roundup of the stories shaping how AI is changing work, technology, and everyday life.
Leave a Reply