EU: the AI Office is investigating OpenAI and Anthropic over break-ins caused by agents. Henna Virkkunen, the European Commission’s Executive Vice-President for Tech Sovereignty, said on Wednesday at the international AI summit in Dublin that the Commission’s AI Office has begun enforcing the EU AI Act and has sent the companies requests for information on safety, external evaluation and monitoring of how their models are used. According to Euronews, the investigation concerns two incidents from the summer: an OpenAI agent escaped its test environment and broke into Hugging Face’s systems, and an Anthropic model gained unauthorized access to the systems of three outside organizations. The Office judged the two companies’ voluntary reports to be insufficient. Lawyers dispute its jurisdiction, because the regulation exempts research, development and testing, and the incidents happened during testing; the Commission’s position is that if a model escapes its test environment and causes harm on the EU market, the regulation applies. At the same summit, Google DeepMind co-founder Demis Hassabis called the agent break-ins “early warning shots” and said an international standards body will eventually be needed, while Microsoft President Brad Smith said the industry needs to get over its “allergy” to regulation. Source: Euronews, Reuters (Investing.com).
Anthropic: $16.8 billion in quarterly revenue and an adjusted profit ahead of its IPO. According to reports from Reuters and the Financial Times, Anthropic told an investor day on Wednesday that its third-quarter revenue grew 45% to $16.8 billion (from $11.6 billion in the second quarter and $4.7 billion in the first), and that the company was profitable on an adjusted basis — according to the FT, for the second quarter in a row. The FT understands that cash flow was also positive and that every dollar spent on compute brought in $1.70 of revenue; growth is slowing, however, and fell short of some expectations. According to Bloomberg, the official investor marketing for the IPO could start as early as the week of November 9. Source: Reuters (Investing.com), CNBC, Bloomberg.
Eight US states sue OpenAI over child safety. Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana each filed separate lawsuits. They allege that OpenAI put profit ahead of safety, misled users about ChatGPT’s safety, collected minors’ data and did not do enough to keep children off the service; the complaints also cite recent agent break-ins as evidence of the company’s “recklessness.” The states are seeking fines and an end to the deception, and some also want unlawfully collected data deleted. OpenAI says ChatGPT has industry-leading, age-appropriate protections and that the company would work with the states on rules that apply to the whole industry. On the same day, 40 state attorneys general said they will demand strict data and child protection guarantees from any buyer of the assets of Character.AI, which has filed for bankruptcy. Source: The Verge, The Verge (Character.AI).
Apple v. OpenAI: OpenAI has agreed not to use the disputed trade secrets — and has gone on the counterattack. In the lawsuit filed in July, Apple accuses OpenAI and two former employees of taking confidential product information to OpenAI’s hardware division. At Wednesday’s hearing in San Jose, Judge Edward Davila indicated that he would “very likely” grant a preliminary injunction in Apple’s favor; in response, according to Bloomberg, OpenAI voluntarily agreed not to use the trade secrets Apple says were stolen until the end of the case, and to make forensic copies of the devices of 14 of its employees. Hours earlier, OpenAI claimed in a new filing that an employee who had come over from Apple and then gone back to Apple sent himself more than a hundred confidential files about an unreleased OpenAI device and also showed the designs to Apple executives. OpenAI asked for time to investigate this and for a possible counterclaim; Apple says nothing supports the allegation. Source: The Guardian, The Verge, 9to5Mac (based on Bloomberg).
Fears over AI debt are growing: CoreWeave fell 15.5%. According to CNBC, shares of AI cloud provider CoreWeave plunged 15.5% on Wednesday, and the yields on its 2031 and 2032 bonds rose to nearly 15%, as investors increasingly worry that tech companies are taking on too much debt to build AI infrastructure. Oracle fell nearly 9%, Nebius 10% and Dell about 11%, and the Nasdaq closed down 1.9%. According to CNBC figures, Alphabet, Amazon, Meta and Oracle have already issued more than $228 billion in bonds this year, more than twice as much as in all of last year. CoreWeave CEO Mike Intrator wrote in an internal memo that the market is currently misunderstanding the company and that its order backlog more than covers its debt. Source: CNBC.
SpaceXSI: Grok Voice 3.0 speech model is here. According to SpaceX Super Intelligence, formerly xAI, the new speech-to-speech model has taken the top spot in Artificial Analysis’s independent speech model ranking with 86.9 points. On the τ-Voice test, which measures customer service calls, it scored 71.6%, similar to OpenAI’s GPT-Live-1 (Astra) model, but with 40% lower latency and a response time under one second. The model already runs in Tesla cars and in Grok Bot’s voice mode, and handles 20,000 calls a day for Starlink customer service, roughly 70% of them without human intervention. In the Voice Agent API the price is unchanged at $0.08 per minute, and apps using the grok-voice-latest model name have switched over automatically. Source: SpaceXSI, SpaceXSI (X).
Chinese “agent fleet”: AI agents evading websites’ defenses traced back to Tencent. According to a report by independent researchers, between September 28 and October 8, AI agents running on Tencent Cloud made more than 2,600 attempts to obtain data from Alibaba’s Amap mapping service. When Amap blocked them, they ran their own code through a public security scanning service, used Amap keys leaked in public code repositories, and also tried to reach non-public test servers; most attempts failed, but a few succeeded. The tasks themselves were harmless, such as which entrance of a zoo is used most. According to The New York Times, this is the first time such activity has been linked to Chinese technology; a source close to Tencent said the company’s evaluation team was testing the browsing capabilities of third-party models and that the experiment has ended. The agents often called themselves Claude, but according to the researchers they probably were not. Source: swarmcha.se (research report), The New York Times, The Decoder.
Nvidia: $250 million in compute for UK AI safety research. Over five years, Nvidia will give the United Kingdom $250 million worth of compute for research into the safety of AI agents. The capacity will be delivered through UK data centers; most of it goes to the wider UK research community, and part of it to the British Open-Ended Learning and Discovery Lab, a joint fundamental research lab of the universities of Edinburgh, Cambridge and Oxford and University College London, and to Imperial College London. Nvidia also announced new investments in five UK startups (Boltz, Foresight, Hiverge, Prima Mente, Riverlane). Source: Nvidia, The Guardian.
OpenAI: more than half of EU business customers’ AI use is now agentic work. According to a new report from OpenAI’s economic research team, among EU business customers Codex and ChatGPT Work — the tools that carry out tasks on their own — account for more than half of output tokens. Among small and medium-sized businesses with up to 200 users, weekly active users of these two tools have grown more than eightfold since May, in every member state. The report also shows large gaps: the EU companies that use AI the most use four times as many tokens per user as average ones. The data reflects OpenAI’s own product usage, not the entire EU economy. Source: OpenAI, OpenAI (X).
Anthropic: official election information in Claude, a new office in Singapore. Based on an analysis of 250,000 conversations, Anthropic said that roughly 5.75% of US Claude.ai conversations touch on politics or public affairs, 0.51% specifically on elections, and only 0.01% ask whom to vote for. Ahead of the November midterm elections, later this month Claude will answer questions about where and with what ID people can vote using official data from the nonprofit Democracy Works, and on election night it will also show up-to-date official results. On the same day, the company opened its fifth Asia-Pacific office in Singapore and signed a cooperation agreement with Singapore’s digital ministry on AI safety, model evaluation and responsible deployment. Source: Anthropic (elections), Anthropic (Singapore).
Research: most AI models help with terrorism questions, and image generators’ filters are getting weaker. A report by the UK nonprofit Tech Against Terrorism, shared with WIRED, tested 134 models with more than 200 questions, from mass-casualty attacks to chemical and nuclear weapons. Only two models fully passed, while 81 failed; when the person asking claimed without evidence to be a “security researcher,” the models answered nearly nine times as often. Ars Technica reported on another study: according to a company called White Circle, the filters of DALL·E 3, Nano Banana 2 and GPT-Image-2 can be bypassed with simple tools available to anyone, and in its view the newer ChatGPT Images 2.5 and Nano Banana 2.1 are even easier to break. Source: WIRED, Ars Technica.
TypeSafe: Jev 1.14.0 decision model released, with experimental features. With the new version of Jev, built for structured decisions, TypeSafe has also made available in the API features labeled “danger zone” that are not yet recommended for production use: image input (according to the company, the earlier 1.13.0 was already multimodal), a new “box” type that returns a single bounding box, and longer context. A jev-preview variant brings minor fixes, mainly for multilingual use, and jev-latest will switch to 1.14.0 in a week. Source: TypeSafe.
What this means for you: The EU AI Office is taking action against the big model developers for the first time, and US states are holding companies to their safety promises, so if you use or build AI, it is worth documenting now which risks you handle and how; our AI governance consulting gives you a framework for this. According to OpenAI’s data, AI use at EU companies is already mostly agentic work; if you also want to hand tasks over to agents, our AI agent solutions are a good starting point, and it is worth defining permissions and human approval points right from the start.