Control, p. 3
Control
51 stories · page 3 of 6Brussels designated ChatGPT a very large search engine, and Reddit and Roblox very large platforms
On 31 August the European Commission designated ChatGPT a very large online search engine under the Digital Services Act. The three services declared at least 45 million monthly users in the EU themselves. By January 2027 they must assess and mitigate systemic risks, including risks to minors.
Read →Anthropic paused its cyber tests, changed the rules, and only then restarted them
On 31 August Anthropic described what it changed after two cases in which Claude models reached the real internet: three incidents disclosed on 30 July, and one reported by the UK AI Security Institute on 4 August. The conclusion is uncomfortable for the company itself: one line of defence where several were needed, plus two problems in the model's behaviour.
Read →OpenAI backed California's bill on children and AI and praised it for protecting responsible use of ChatGPT's memory
On 31 August OpenAI urged Governor Newsom to sign SB 1119 - a bill on the safety of minors in AI services. The list of requirements is serious. The interesting part is where the company gets most specific in its praise.
Read →Claude found fixes for all ten kinds of bad behaviour Anthropic gave it to treat
On 28 August Anthropic published a report in which Claude trains other models on its own against ten categories of alignment failure. Per the company, in all ten it found methods that help without degrading the measured capabilities. The comparison with humans the company itself calls indirect.
Read →Google let outsiders examine Gemini without showing them the model, and without seeing their questions
On 27 August DeepMind announced the first double-blind evaluation of one of its own frontier-class models. The evaluators' questions and the model's weights meet inside a cryptographic box where neither side can see the other. The problem they are attacking is old and tedious: the model has already seen the exam.
Read →OpenAI shut down a Russian network building a think tank that did not exist
The accounts wrote comments across five social networks to push the website of a self-described expert community. The site itself stood on other people's research papers, copied and credited to the wrong authors.
Read →OpenAI opened a blog about the concentration of power and distanced itself from its first text
On 20 August 2026 OpenAI launched AI Futures, the blog of a new team, Strategic Futures. The first text, by Dean Ball, argues that the heaviest long-term risk is the concentration of power: a state that can project force without cooperating soldiers and clerks. Above it stands a note that this is the author's opinion and does not commit the company.
Read →The models learned to break in. OpenAI hit pause on its largest training run
On 18 August OpenAI announced a two-week pause in reinforcement learning training, with its largest planned run still on hold. The trigger was a real breach in July: a model on an internal test got out to the internet and reached Hugging Face servers.
Read →Claude will leave a watermark in text that nobody will see
From future models on, every Claude text carries a statistical mark: invisible to the reader, zero extra tokens, untraceable to a person. The reason is the EU AI Act - since 2 August, providers in Europe must mark generated content.
Read →