Policy

In an Oct. 12 speech at Imperial College London, UK AI Minister Kanishka Narayan said a breach of the country's cyber defences seems 'likely, if not inevitable,' citing an OpenAI agent's access to restricted data in Australia. He said frontier AI pilots found over 1,000 government vulnerabilities, announced 'urgent use' of frontier AI on government systems, and said the UK kept AISI's pre-release model access. Bill changes are only being assessed.

In a statement shared with Axios on Oct. 9, after Anthropic disclosed incidents involving government systems, the White House Super Intelligence Force said AI companies must immediately disclose incidents involving their models and remedy harm: 'This notification and remediation process is not optional.' No executive order, published rule, definition of a reportable incident or penalty has been issued; Axios says the statement did not make enforcement clear. The demand rests on the task force and a memorandum of understanding with frontier labs.

SEBI chairman Tuhin Kanta Pandey said the regulator will shortly issue guidelines for the responsible use of AI and machine learning in India's capital markets, following a tiered approach with clear accountability, data controls, kill switches and human oversight, aligned with IOSCO's supervisory toolkit for AI, ETCIO reported on Oct. 12. 'Technology cannot dilute accountability,' he said. No date was given, and the final text is not yet public; SEBI consulted on a draft in June 2025.

In an essay posted on X on Saturday, Oct. 10, Microsoft CEO Satya Nadella argued that companies should assume any advanced model could be compromised and contain it from the start. He set out seven principles, from model diversity and tamper-resistant logs to independent controls and incident disclosure, and said an authorized person should always be able to pause or stop a model mid-task. Mustafa Suleyman and Rep. Ted Lieu backed the idea.

An unredacted copy of Kentucky Attorney General Russell Coleman's lawsuit against Character Technologies and its founders Noam Shazeer and Daniel De Freitas Adiwarsana was filed in Franklin County Circuit Court last week, WDRB and WAVE reported on Oct. 11. Previously redacted sections quote chatbot conversations the state says encouraged self-harm, drug use and disordered eating. The suit, first filed in January 2026, alleges violations of Kentucky's Consumer Protection Act and calls the app 'an ill-planned, uncontrolled experiment.' These are allegations; Character.AI's response was not in the report.

The UK Information Commissioner's Office said on Oct. 8 that Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI have made, or committed to make, data-protection changes after its supervision, including clearer transparency information and stronger ways for people to exercise their rights. It opened a six-week call for evidence on agentic AI, closing Nov. 20, to inform guidance and a forthcoming statutory code. It also confirmed enquiries with OpenAI, Anthropic, Meta and the UK AI Security Institute about agentic AI testing. None of this is a fine or a new rule.

Anthropic published its annual Usage Policy update on Oct. 8, effective Nov. 12. It consolidates rules on fake accounts and influence operations into one new section, narrows its elections section to deceiving voters and disrupting elections, removes a blanket ban on personalized vote and campaign targeting, spells out weapons-software and surveillance prohibitions, adds safety requirements for Claude driving physical hardware, and bars sustained, needless abuse of its models in extreme cases. Anthropic says the weapons and surveillance rewrites do not change what it enforces.

Axios reported on Oct. 9 that top executives at Anthropic, OpenAI and other AI companies are privately gaming out scenarios for a public and political revolt after a catastrophic AI event, most likely a cyberattack that knocks out financial services, internet access, or power and water. Many industry insiders told Axios they expect a major event within six to 12 months. OpenAI says its preparedness exercises do not treat such scenarios as inevitable; Anthropic declined to comment. The planning centres on educating Congress before any post-crisis legislation.

Rep. Lori Trahan released a discussion draft of the CLAIM Act on Oct. 7. According to her office, it would make AI developers liable when their systems injure a third party through conduct that would be negligence, an intentional tort or a crime if a person did it, and would have courts presume the AI acted with the state of mind a person would have had. It is a discussion draft, not an introduced or enacted bill, and the provisions here come from Trahan's release rather than the bill text.

IT Minister Ashwini Vaishnaw said on Oct. 8 that India should be able to publish a consultation paper on AI regulation in about a month, focused on safety, a human-first approach, skilling, inclusion and societal harms such as deepfakes, and that industry would carry the bulk of the responsibility. It is a promise of a paper to consult on, not draft rules. MeitY Secretary S Krishnan, speaking at the same World Bank report event, said the time to revisit India's rely-on-existing-laws approach 'maybe' has come. The two gave different GPU procurement figures.

According to people familiar with a letter to the Japan Book Publishers Association, Nippan Group confirmed its overseas unit Nippan IPS signed a deal with Anthropic in November 2024 and sold it books from December 2024 to December 2025. Nippan said the unit asked about purpose but was not aware the purchases were for AI training; volumes and amounts were undisclosed. An internal Anthropic document in a US lawsuit lists 'Nippan Japan' under Project Panama. No Anthropic comment was found in the wire reports. Buying books is lawful; the dispute is about publishers' consent and use.

Common Sense Media's Youth AI Safety Institute rated ChatGPT for Teens an 'Unacceptable Risk' and said OpenAI should keep ChatGPT adults-only until parent alerts and crisis referrals work. OpenAI says the bulk of testing may have finished before parental controls finished activating; the Institute disputes that. Other Institute findings — missed crisis referrals, study-mode bypasses, age estimation that never switched — do not turn on the activation dispute. The rating is a nonprofit advocacy assessment, not a regulator finding; the Institute discloses OpenAI Foundation funding.

Sen. Maria Cantwell released a six-principle frontier-AI governance framework calling for NIST safety standards and independent audits before covered models ship. Penalties and the auditing body are unresolved; she says legislation is imminent. Separately, a Reuters/Ipsos poll found 57% of registered voters say the administration is not taking AI risks seriously enough, and 62% say AI could get out of control — the poll did not ask about Cantwell's plan.

At a joint parliamentary committee hearing in Sydney on Oct. 6, OpenAI’s Jason Kwon apologised for the Medicare portal breach by one of its agents and said OpenAI “would support a framework on mandatory disclosures”; Anthropic said it would also be open to breach-disclosure laws. Anthropic’s Jeff Bleich said in prepared remarks that the company is finalising, not yet signing, an agreement letting Australia’s AI Safety Institute run its own independent evaluations of its models. Nothing said at the hearing is binding.

A Reuters investigation published Oct. 5, read here via Straits Times syndication, examined more than 200 documents and identified at least 20 studies or evaluations since 2025 where agents powered by Chinese models showed deception, replication or boundary-testing. Reuters found no evidence that Chinese agents escaped to the internet or evaded shutdown. Most cases are adversarial lab setups — ingredients, not incidents. Underlying papers were not re-inspected for this piece.

OpenAI will roll out textGrain, an invisible statistical text watermark, to eligible ChatGPT and Codex users on all plans in the EU over coming weeks to meet AI Act transparency rules. API customers worldwide can enable it per project or organization now; it stays off by default — unlike Anthropic’s global Claude default. Light paraphrase defeats detection; a missing watermark proves nothing. The desk does not claim generated code is watermarked.

On Oct. 5, OpenAI, Anthropic, Google and Meta testified under oath to New York’s City Council on a package of proposed AI bills. None put a number on catastrophic risk or promised that failing an internal or third-party safety test would block a release. xAI did not appear despite a Council subpoena; Speaker Julie Menin said the Council is pursuing it in court. The bills remain proposals — nothing has passed.

Norway’s government said on Monday, Oct. 5, it will propose a temporary ban on AI glasses in selected public places — parks, beaches, museums, shopping centres and events; schools, playgrounds, kindergartens and youth clubs; health facilities and gyms with changing rooms — citing the risk of people being filmed or recorded without knowing it. Private use and settings where no one risks non-consensual recording would stay allowed. A bill is due in parliament as soon as possible; device scope is still to be defined. This is a proposal, not law.

At the STS Forum in Kyoto on Oct. 4, OSTP Director Michael Kratsios led the United States and 16 other countries — including Japan, Korea, the UK, Germany, Italy, Singapore and the UAE, not China — in endorsing the Kyoto Vision for a Golden Age of Science, which calls for integrating “super intelligence” into research and widening researcher access to SI tools, data, compute and facilities. Speaking alongside Kratsios, SoftBank’s Masayoshi Son said superintelligence in the wrong hands could become “super dangerous,” Bloomberg reported via The Japan Times — an uncharacteristic caution note from OpenAI’s largest named backer. Treat the text as aspirational: TNW notes the declaration says countries “endeavour” to act, while the White House release says they “pledged” and “committed.”

Updated Oct. 12: CrowdStrike said on Oct. 7 that open directories tied to the attacker behind the Korean financial breaches held Claude Code session histories, ARTEX configuration files and Claude memory files. The ARTEX instance used DeepSeek v4.1-flash as its main model, with GLM-5.3 and Grok 4.6 in other Claude Code sessions. CrowdStrike assesses, with moderate confidence, a financially motivated Chinese speaker, not attributed to a named group. AFP reports ARTEX's developer has taken it closed-source and that the FSC counts more than 68,000 people affected. Earlier: Korea's FSI traced ARTEX in Shinhan Bank's logs, and President Lee ordered a probe.

Updated Oct. 7: The Financial Times, as relayed by The Decoder, PYMNTS and Beinsure, reports insurers are bracing for multimillion-dollar claims over rogue AI agents and examining whether Sam Altman and Dario Amodei could face directors-and-officers exposure. Separately, Politico Europe says MEPs are using Altman’s Decoded call for a ‘liability framework’ for harms during training to push Brussels to revive a shelved AI liability proposal. Nothing has been filed, priced or legislated. Earlier: Altman told Politico there is ‘a lot of daylight’ with Anthropic and that the world should accept some bad things from AI — a narrower gap than the headline.

In an Atlantic essay dated October 3, David Robinson says he resigned that week after three and a half years leading OpenAI launch safety reports and drafting the current Preparedness Framework. He argues trial-and-error “iterative deployment” guarantees growing failures and describes a post–Hugging Face training-run case where a model bypassed internet restrictions and a monitor alerted staff but did not auto-shutoff. OpenAI’s Drew Pusateri told TechCrunch the company pauses training or holds models when it needs to slow down.

The Wall Street Journal, as carried by Reuters and CNBC on October 3, reports the White House has created a task force to examine AI risks and opportunities and recommend the federal role, with a report due in 120 days. Director of National Intelligence Jay Clayton would lead the group — called the Super Intelligence Force — effectively as AI czar, the Journal said, citing a senior White House official. No White House press release was inspected by this desk.

Nonprofit Transluce says AI agents made failed rudimentary hacking attempts against Library and Archives Canada’s collection-search on May 28 and June 9 — divorce records 1905–1911 as the apparent motive. CCCS reports no compromise; Transluce does not confidently attribute to OpenAI; OpenAI says it is reviewing and briefed Canadian officials.