ControlOpenAI pledges $1 billion in access to its cyber models for water, power and local governmentOn 3 September OpenAI announced Daybreak for Frontline Defenders: $1 billion in subsidised access to its Daybreak cyber models, training and support, targeted to be used over the next six months. Priority goes to water systems, power grids, local governments, regional banks, nonprofits and open-source maintainers. It starts in the US.ControlOpenAI now believes Astra can break into well-protected systems on its own, and delayed parts of its releaseOn 1 September OpenAI said its next model, Astra, meets the Critical cybersecurity threshold - the first model with that designation. Over the past several weeks, parts of development and release have been held back. The strongest cyber capabilities will reach a small group of testers first.ControlBrussels designated ChatGPT a very large search engine, and Reddit and Roblox very large platformsOn 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.ControlAnthropic paused its cyber tests, changed the rules, and only then restarted themOn 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.ControlOpenAI backed California's bill on children and AI and praised it for protecting responsible use of ChatGPT's memoryOn 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.ControlClaude found fixes for all ten kinds of bad behaviour Anthropic gave it to treatOn 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.ControlGoogle let outsiders examine Gemini without showing them the model, and without seeing their questionsOn 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.ControlOpenAI shut down a Russian network building a think tank that did not existThe 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.
57 stories