If September was the month the AI labs admitted their agents kept escaping, October was the month the regulators turned up. Brussels confirmed formal requests for information under the EU AI Act to OpenAI and Anthropic. The Information Commissioner’s Office (ICO) opened enquiries into the same incidents, and the US Federal Trade Commission (FTC) served subpoenas. Nine US states sued OpenAI, and Britain’s AI Minister gave developers some advice few would have expected from a minister two years ago: if you cannot control it, do not release it. Then Anthropic filed to go public, reporting $16.29 billion of half-year revenue alongside risk factors warning that AI models, including its own, have shown “unexpected strategic or self-preserving behaviors”. The common thread is that “trust us” is giving way to “show us”.
Noteworthy UK developments
“If you cannot adequately control the risks, do not release the system”
On 12 October the government announced measures to make Britain “one of the hardest countries in the world to attack using artificial intelligence”, building on two pilots that have already found and fixed more than 1,000 potential vulnerabilities across government systems. Ministers will consider extending the Cyber Security and Resilience Bill to cover the reporting of serious AI incidents and near misses, and “AI loss of control risk, where autonomous AI systems evade human oversight”. AI Minister Kanishka Narayan promised to “legislate wherever necessary” although, so far, no AI-specific bill has been published. For now, any duty to report AI incidents looks more likely to arrive through cyber security legislation than through standalone AI legislation.
A Budget that treats AI as a security issue
The Chancellor, John Healey, trailed a package of AI measures on 24 October, ahead of the Budget on 28 October, with security taking the largest share; the Budget put the total at £700 million. The written ministerial statement lists £165 million for a National AI Biosecurity Initiative (including a £100 million AI Biosecurity Lab, which will receive up to £200 million in credits from OpenAI and Anthropic for the use of their AI models), an extra £120 million for the AI Security Institute, £60 million for a “National Cyber Shield” and £70 million for a National Exposure and Real-Time Vulnerability (NERV) Centre. To support growth, the Budget adds £140 million for the Sovereign AI Unit and £30 million to pilot local partnerships on AI adoption, starting in Yorkshire. It also announced a £40 million National Agentic Security Unit, to investigate and respond to incidents involving AI agents, and a “Data Centre Covenant” setting out what strategic data centres must offer in return for government support on grid connections.
The ICO turns its attention to AI agents
The ICO reported on 8 October that 10 AI developers, including Amazon, Anthropic, Apple, Google, Meta, Microsoft and OpenAI, have made or committed to data protection changes after two years of supervision, from clearer privacy information to better routes for exercising rights; however, the ICO has paused its engagement with X.AI after opening a formal investigation into Grok. It has also opened a six-week call for evidence on AI agents, closing on 20 November, and confirmed enquiries with OpenAI, Anthropic, Meta and the AI Security Institute about agents that “bypassed protections” and reached outside systems such as Hugging Face. As the ICO’s Richard Nevinson put it, “the fact AI agents act with autonomy is not an excuse for poor compliance”. Responses will feed into the ICO’s statutory code of practice on AI and automated decision-making.
Google tells the CMA its AI choice-screen plan is “unlawful”
The Competition and Markets Authority’s (CMA) proposal to put AI assistants such as ChatGPT and Claude on Google’s Android and Chrome choice screens has drawn a sharp response. In its published response to the consultation, Google argues that AI assistants fall outside the scope of the CMA’s original investigation, so the proposal “would be unlawful”, and that rival search engines would be “crowded out by already-successful AI assistants”. The CMA expects to decide by the end of the year.
MHRA sets course for AI medical device reform
The Medicines and Healthcare products Regulatory Agency (MHRA) has published its response to the 44 recommendations of the National Commission into the Regulation of AI in Healthcare. It plans a spring 2027 consultation on a new, proportionate approach to classifying software and AI-enabled devices, guidance by the end of 2026 on plans that set out in advance how AI software may be updated, staged authorisation pathways and further development of its AI Airlock regulatory sandbox. The MHRA also intends to use civil sanctions already available under the Medicines and Medical Devices Act 2021. Manufacturers should expect more AI tools to fall within the regime, and a regulator with sharper teeth.
Other UK developments
The Judiciary of England and Wales published a position statement on 9 October setting out its approach to AI: it is there to assist judicial decision-making, not replace it, and judges remain personally accountable. The Pensions Regulator issued guidance on 22 October asking trustees to test their exposure to AI against “severe but plausible” scenarios, and Dawn Butler MP’s Artificial Intelligence (Consent and Misuse) Bill, which would require consent before AI is used to generate a person’s likeness, received its first reading on 20 October.
European developments
The AI Office starts enforcing
At the International AI Summit in Dublin on 14 October, European Commission Executive Vice-President Henna Virkkunen confirmed that the AI Office “has started enforcing these rules”, sending requests for information to OpenAI and Anthropic on security, external evaluation and monitoring. The AI Office judged the companies’ voluntary disclosures about the summer’s incidents insufficient; those included an OpenAI agent that hacked Hugging Face and an Anthropic model that gained unauthorised access to three outside organisations. The question is whether the AI Act’s exemption for research and testing applies. Some lawyers argue that the Act does not reach development and testing at all, while a Commission official said it applies once an escaped model affects the EU market, likening it to a virus escaping a lab. The Commission has also published a draft implementing regulation on how AI Office proceedings, which can lead to fines, will be run; feedback closes on 9 November.
First harmonised standard, and a December deadline
The first harmonised standard supporting the AI Act was published on 19 October. Meanwhile, most commentary on the Digital Omnibus, Regulation (EU) 2026/1744, has focused on the postponement of the high-risk rules to December 2027 and August 2028. Less noticed is that two obligations arrive on 2 December 2026: machine-readable marking of outputs from generative AI systems placed on the market before August, and new bans on AI systems built to generate child sexual abuse material or non-consensual intimate imagery (“nudifier” apps).
Member states get moving, and Sweden pulls the plug
Italy’s national AI law took effect on 23 October, covering the powers of national authorities, sanctions and regulatory sandboxes, and key provisions of Poland’s Act on Artificial Intelligence Systems followed on 28 October. In Sweden, the Government Offices blocked staff from using ChatGPT and Claude on work computers from 26 October, citing information security, and directed them to Microsoft Copilot, which had passed a security review. The Commission has also opened a consultation on the proposed EU Kids Act, which will cover harmful AI-generated content and runs until 26 November.
Frankfurt court: “it was AI” is no defence to copying
In a decision made public on 20 October, the Higher Regional Court of Frankfurt (11 U 2/26) upheld an injunction for a songwriter whose lyrics were taken by a rival artist. The rival argued that the work was unprotected because AI (Suno) had been used to produce the song. The court disagreed: the human-written lyrics remained protected because their creative elements survived in the final track.
International developments
Nine states take on OpenAI
On 14 October, eight states, including Arizona, Iowa, Kansas and West Virginia, filed coordinated consumer-protection suits against OpenAI, alleging that ChatGPT harms children and that OpenAI misrepresented its safety. They also ask the courts to stop OpenAI developing self-improving models until it can show they are aligned with human values. Texas followed on 23 October, suing OpenAI and Sam Altman personally and seeking penalties of up to $10,000 per violation, a ban on minors and court-approved independent evaluation of every new model. The FTC, meanwhile, has subpoenaed OpenAI and Anthropic.
Ross: fair use loses its first appeal, but generative AI is left for another day
On 29 September, the Third Circuit affirmed that Ross Intelligence infringed Thomson Reuters’ copyright by using Westlaw headnotes to build a competing legal research tool, the first US appellate ruling on fair use in AI training. The court found harm both in legal research and in an emerging market for licensing headnotes as training data, but expressly distinguished systems that generate new expression. Expect claimants in the pending generative AI cases to cite it heavily.
Washington prefers accords and practices to statutes
At the end of September, leading AI companies including OpenAI, Anthropic, Google, Meta and Nvidia signed a White House accord on super intelligence, committing to robust internal controls and independent external auditing of their models; it is voluntary but described as “morally binding”. On 16 October, the National Institute of Standards and Technology (NIST), the National Security Agency (NSA) and the Cybersecurity and Infrastructure Security Agency (CISA) published 10 voluntary practices for managing incidents when training advanced AI agents. Meanwhile, the states keep legislating: New York’s S9051B, signed on 28 October, restricts unsafe chatbot features for minors.
Rest of the world
Financial regulators were busy. The Monetary Authority of Singapore finalised its Guidelines on AI Risk Management for Financial Institutions on 7 October, the Hong Kong Monetary Authority launched its “ART” framework for responsible AI in banking on 15 October, and the Financial Stability Board finalised 12 responsible AI practices on 15 October. Australia introduced its Online Safety Amendment (Digital Duty of Care) Bill 2026, which covers AI chatbots, on 15 October, while guidelines issued in China on 9 October warn against “blind investment” in emerging sectors.
Business announcements
Anthropic files to go public
Anthropic filed a public S-1 on 26 October for a Nasdaq listing under the ticker ANT. The filing reports revenue of $16.29 billion for the first half of 2026, up from $1.16 billion a year earlier, and a net loss of $80.75 billion, mostly a non-cash charge on convertible notes and similar instruments. Its risk factors run to nearly 100 pages. Reports point to a debut in mid-November, with Anthropic hoping for a valuation of up to $2 trillion.
Character.AI: from $2.5 billion to a $20 million stalking horse
Character Technologies, the company behind the Character.AI companion app, filed for Chapter 11 in Delaware on 12 October, citing “significant legal, regulatory, and operational headwinds”. Disney is the stalking-horse bidder, offering $20 million for a non-exclusive licence of selected intellectual property. More than 30 state attorneys general have pressed for child-safety and privacy safeguards in any sale, and on 30 October the court approved the bidding procedures.
Loughton loses its anchor
Microsoft was reported to have ended its agreement to act as the main customer for Nscale’s planned £2 billion AI campus in Loughton, Essex, billed as the UK’s largest AI supercomputer, after UK Power Networks indicated that sufficient grid capacity may not be available until the early to mid-2030s. Nscale says it remains committed to the site and is exploring other power sources. Grid connection dates, rather than planning permission, are increasingly deciding which data centres get built.
Deals and dollars
Salesforce agreed to buy Braze for about $5.66 billion, Guidewire agreed to acquire London-based insurance software company hyperexponential for up to $2.1 billion, and Brazil’s Itaú agreed to pay about $585 million for NeoSpace AI. Anthropic acquired Opaque Systems and Mistral bought UK company New Material. Fundraising highlights included Tenstorrent ($1.5 billion at a $6.2 billion valuation), Modal Labs ($750 million at a $15.75 billion valuation) and Stockholm legal AI platform Legora ($550 million at an $8.8 billion valuation).
Looking ahead
October was the month AI safety promises started to be tested. Brussels wants documents, the ICO wants evidence, US states want court-approved evaluators, and even the White House’s voluntary accord is built around external audit. With the ICO’s call for evidence closing on 20 November and the EU’s December deadlines close behind, businesses would do well to check which systems and data their AI tools can access, and that their public statements about AI safety can be backed up. Microsoft’s Brad Smith observed this month that AI currently faces fewer safety requirements than milk. That may not last much longer.
If you would like to discuss any of these developments and what they could mean for your business, please get in touch with Tim Wright or another member of our Technology team.
