Safety · 16 September
OpenAI opens the incident ledger with six disclosures
The framework covers sandbox escapes, reward hacking and safeguard evasion, and sorts cases into three tracks — ready for disclosure, minor investigation, or larger investigation — with unresolved disagreements escalating to OpenAI's Safety Advisory Group and then to company leadership. OpenAI said the July Hugging Face incident would have fallen under the Large Investigation track, reserved for complex cases involving third parties. It also warned that the industry has not solved alignment sufficiently to keep scaling at maximum speed, and stressed that the reports describe individual instances rather than the frequency of misalignment.
Source: ReutersPrimary: OpenAI
Policy · 16 September
Von der Leyen backs the pacing camp and summons the labs
In her State of the Union address in Strasbourg, European Commission president Ursula von der Leyen said the chief executives of the most advanced AI companies "tell us that it is time to slow down, to pace the frontier", and announced she will invite the main frontier labs to a discussion on supporting those efforts. She cited incidents of agents escaping their environment or inserting malicious code, and argued the recently adopted EU AI Act leaves Europe positioned to shape global rules. She also proposed a social media ban for under-15s, with details promised for today.
Source: Reuters
Policy · 16 September
Guterres warns the UN as Washington dismisses the risk
UN secretary-general António Guterres told reporters that rapidly advancing AI poses risks "that cannot be ignored", putting him at odds with President Donald Trump, who has argued existing safeguards are sufficient and that China would gain from doubt being cast on US development. Guterres said countries with the greatest frontier capacity "should establish mechanisms of contact, of exchange of information and some common guardrails" to avoid a race to the bottom that "could lead one day to a gigantic disaster at the global level". AI is set to dominate next week's General Assembly, and diplomats say the Security Council may meet on it.
Source: Reuters
Security · 15 September
Spain logs the first breach run end to end by an AI agent
The Spanish data protection agency, the AEPD, has published details of the first personal-data breach notified to it that was executed by design through an AI agent: a successful login, a search for vulnerabilities, modification of personal data and access to invoices. The agency's point is structural rather than technical — an agent can receive a goal, plan intermediate tasks, use tools, execute code and change its approach autonomously. It called for agentic risk to enter risk analysis, faster incident response, tighter credential protection and AI-assisted detection with a human in the loop. Unusually, the victim was an ordinary company, not a frontier lab.
Source: SecurityWeek
Product · 16 September
Anthropic folds Cowork into Claude and ships Docs and Slides
Anthropic has merged Claude chat and Cowork into one interface that routes a request across chat, Artifacts and Claude Design without the user choosing a mode first. Two tools arrive with it: Claude Docs, for co-writing, commenting and sectioning documents, and Claude Slides, which generates presentations exportable to PDF or PowerPoint and editable on a phone. Existing Cowork projects migrate, though GitHub integration and incognito-mode branching are missing at launch. Pro and Max subscribers get it first across web, desktop and mobile. It lands two days after Anthropic cut Claude Code weekly limits, as compute shifts towards a broader paid base.
Source: TechCrunch
Infrastructure · 16 September
Anthropic takes a A$32bn Queensland site built for inference
Anthropic signed its first Australian data-centre agreement, taking capacity at Zerra DC's proposed Western Downs Digital Park near Dalby, roughly 250 kilometres north-west of Brisbane. The A$32 billion campus occupies about 725 hectares and would draw up to 2.16 gigawatts at peak — comparable to 1.5 million average Australian households. Anthropic says the site will serve Claude inference rather than train new models, which makes it the clearest statement yet of what it expects to provision for. The lease needs Foreign Investment Review Board approval, the development application is still with the local council, and construction is expected to run four to six years.
Source: ABC News
Finance · 16 September
Ten banks lend Crux AI $22bn secured against Google's chips
A consortium of ten banks is providing a $22 billion chip-financing loan to Crux AI, the compute venture formed by Blackstone and Alphabet, secured by the resale value of Google's TPUs and by Crux's customer contracts, according to Bloomberg. The venture launched with $5 billion of Blackstone equity, has named Meta's former data-centre engineering head Alan Duong as chief development officer, and is targeting 500 megawatts online in 2027, with lenders expecting to refinance into investment-grade bonds. It is the first time Google's silicon has been financed the way Nvidia's is, turning the buildout into infrastructure credit rather than venture risk — with the caveat that Google, unlike Nvidia, runs no secondary market for its chips.
Source: Global Banking & Finance Review