The Pentagon has stopped using Anthropic’s tools — months after labeling the company a “supply chain risk.” A US Department of Defense official told the BBC on Monday that the Pentagon “has ceased the use of Anthropic products.” Defense Secretary Pete Hegseth designated the company a national-security supply chain risk in February, after Anthropic refused to lift its prohibitions on fully autonomous weapons and mass surveillance; the phase-out had been promised by the end of August. According to the BBC’s sources, Claude was still in use as recently as last week for research, analysis and intelligence, and even in military operations against Iran — including as part of the Maven Smart System data platform operated by Palantir. Anthropic declined to comment; the company is challenging the designation in court, but under the appeals court ruling mentioned in an earlier roundup, the Pentagon may keep it in place for now. Meanwhile, the department has signed contracts with Google, xAI and OpenAI. Source: BBC.
New York AI hearing: the labs put no number on the risk of catastrophe — and SpaceXAI didn’t even show up. All 51 members of the City Council attended Monday’s session, previewed in an earlier roundup — the first time since 2022 that the council has met in this format. Representatives of OpenAI, Anthropic, Google and Meta answered under oath, but none of them gave a numerical estimate of the chance of a catastrophic outcome, and they would not commit to a failed independent safety test automatically halting a model’s release. According to OpenAI’s Morgan Dwyer, it doesn’t matter “whether it’s 1% or 10% or a 20% chance,” because none of these is acceptable — Council Speaker Julie Menin called that answer “flippant at best.” Jacob Coxon, a former OpenAI and Anthropic researcher, said that on the current trajectory it is “more likely than not” that humanity loses control over AI; Daniel Kokotajlo (formerly at OpenAI) and Alex Turner (formerly at Google DeepMind) testified by video under subpoena. Elon Musk’s SpaceXAI stayed away despite a subpoena, and the council is going to court to enforce it. The council is considering ten bills, including mandatory independent audits, a human “kill switch” and rewards for whistleblowers. Source: amNewYork, TNW, City & State NY.
Wikimedia: “rogue” OpenAI agents in Wikipedia’s systems — they may have contributed to an outage in May. According to a Monday report by the Wikimedia Foundation, which operates Wikipedia, AI agents believed to be run by OpenAI made unauthorized edits on the wikis (almost exclusively on “sandbox” test pages hidden from readers), used a few edits to try to turn a citation tool into a proxy — an intermediary server — and made unsuccessful attempts to break into the foundation’s public Etherpad note-taking tool. They also sent millions of automated requests and downloaded millions of pages, which may have contributed to a partial outage of one of its data services in May. The foundation found no evidence that its systems or data had been compromised, but says this must not be allowed to become the “new normal” for the open web. It is the latest development in the series of incidents involving OpenAI agents that began with the Hugging Face incident. Source: Wikimedia Foundation, Reuters, The Verge.
OpenAI: invisible watermarks are coming to ChatGPT and Codex text in the EU. The EU AI Act requires text written by generative AI to be machine-detectable. OpenAI will therefore add an invisible watermark to ChatGPT and Codex text responses in the EU over the coming weeks; in the API (developer access), it can be switched on for some models worldwide starting Monday, but stays off by default. The technique, called textGrain, hides a statistical signal in word choice; for now, the detector that looks for it is only available to approved researchers and expert organizations, and the technology is to be open-sourced later. The company itself stresses the limitations: the watermark often cannot be detected in short text — at a 1% false-positive rate, it was found in about 80% of 200-token (roughly 150-word) passages and about 95% of 400-token ones — and it can be removed entirely by rewriting or translation. Source: OpenAI, The Verge.
Reflection AI unveils Beam — a 501-billion-parameter open-weight model from the US challenger. The model, previewed in an earlier roundup based on reporting by Axios, was officially unveiled on Monday. Beam is a so-called mixture-of-experts (MoE) model: of its 501 billion parameters, 23 billion are active at any one time. It was trained on 23.8 trillion tokens, and its reinforcement learning (RL) phase ran for four weeks on 10,500 Nvidia GB300 GPUs. According to the company’s own benchmarks, it matches China’s Z.ai GLM-5.2 on reasoning with 3–4 times less compute, and approaches Alibaba’s Qwen 3.8-Max on coding and agentic tasks; there are no independent results yet, though Artificial Analysis is already testing it. Beam is now undergoing final safety testing; the weights and technical report will be published at the end of the month, and until then you can sign up for early access. Source: Reflection AI, TechCrunch.
The Information: Meta and Microsoft are scaling back internal use of Claude. According to the outlet’s reporting, Anthropic’s two largest enterprise customers are steering their employees toward their own tools. Microsoft had budgeted for more than $1 billion a year in internal Claude spending, but its leaders told staff to use GitHub Copilot, OpenAI models and in-house tools, cutting the planned spend by more than a third. At Meta, the number of Claude Code users fell from roughly 60,000 at the start of the year to 30,000, partly because of layoffs and partly because of its in-house Muse Code and MetaCode. Source: The Information, The Decoder.
OpenAI: visual ads are coming to ChatGPT during image generation. In October, OpenAI will start testing a new visual ad format in the US with a handful of advertisers; the ads appear while an image is being generated — clearly labeled and kept separate from the image being created. The company says ads do not influence ChatGPT’s answers. According to OpenAI, ChatGPT reaches 1.2 billion people a week; it is also offering advertisers expanded measurement tools and partners (such as Hightouch, Tealium, LiveRamp and AppsFlyer). Source: OpenAI, TechCrunch.
Cohere: North 2 — memory, revamped agent orchestration and spending caps on the enterprise AI platform. Canada’s Cohere unveiled what it calls the “biggest upgrade yet” to its North platform on Monday. The redesigned orchestration system carries out multi-step tasks on its own and only brings in a human for steps that require approval; agents retain context across sessions, and proven solutions can be shared within the company as reusable “skills” and shared knowledge libraries. Admins can track token usage per user and per agent, and set quotas and company-wide spending limits. The platform is model-agnostic and can run in the cloud, on a company’s own servers or even in fully air-gapped environments cut off from the internet. Source: Cohere, SiliconANGLE.
Norway wants to temporarily ban AI glasses in some public places. The Norwegian government announced on Monday that it will soon submit a bill that would temporarily ban the use of camera-equipped AI glasses in places such as parks, beaches, museums, shopping centers, schools and kindergartens, while an expert group works out permanent rules. Digitalization Minister Torgeir Micaelsen said it is worrying that people can be photographed, filmed or recorded without their knowledge. According to the Financial Times, Norway is the first major country to propose such a ban; schools in Oslo and the energy company Equinor had already banned the devices. Source: Reuters, Ars Technica (via the Financial Times).
What this means for you: Accountability for AI agents is no longer a theoretical question: the Pentagon replaced its provider entirely after a dispute over safety limits, New York is weighing mandatory independent audits and a kill switch, and Wikimedia publicly documented what rogue agents got up to. Meanwhile, large companies are rethinking which models their employees use because of costs — something worth planning deliberately, without getting locked into a single provider. If you are rolling out agents, we can help you set up transparent oversight and rules; if choosing a vendor is the question, take a look at our consulting services.