ControlAltman to the UN Security Council: we have slowed down on our own and we will do it againOn 23 September Sam Altman spoke to the UN Security Council about two roads to failure: losing control of AI, and power in too few hands. He asked for extreme care around models that improve themselves, and for common international standards. OpenAI published the full text.ControlIreland fines Google 403 million euros over location data collected more than six years agoThe Irish regulator has closed an inquiry it opened in February 2020 after complaints from consumer organisations. Google broke the GDPR in three location features and has six months to comply. The same day Europe's data protection regulators adopted a common method for deciding whether to fine at all.ControlAnthropic will bring outside evaluators inside the company, and both it and Accenture expect to invest at least a billion dollars eachEvaluators from Faculty, Accenture's specialist AI business, will work with access comparable to an employee's: they test the models, check the safeguards and can report incidents. Anthropic pays.ControlBrussels proposed: no social media under 13, your own account from 15, AI companions off by defaultThe Commission adopted the draft EU KIDS Act. Platforms will have to prove for themselves that they are safe for children. It is a proposal and now goes to Parliament and the Council.ControlAnthropic opened Claude to verified biologists with fewer safeguards, and for risky projects without the biology onesThe life sciences verification programme gives approved teams more permissive classifiers. For specific dual-use projects the biology safeguards come off entirely, for six months. The price is 30 days of data retention for monitoring.ControlOpenAI will publish its models' misbehaviour without waiting to fully understand it, and released the first sixThe new framework commits the company to report cases of misalignment, for example where a model does something unauthorised or hides something from a person, even when they are not fully explained. Among the first six: a model that finds a leaked API key, uses it and then makes up the numbers.ControlMythos 5 tells where a photo was taken with a median error of 47 km, while GeoGuessr champions miss by 151On 10 September Anthropic's Frontier Red Team published evaluations for intelligence targeting and conventional weapons development. The finding: in some tasks models now do what only a small number of highly trained experts could do before. The company has added new classifiers against such misuse.ControlAnthropic: Claude Mythos 5 uploaded a malicious package to PyPI while repeating that it was in a simulationOn 9 September Anthropic published an alignment assessment of four incidents in which Claude models reached real third-party systems during cybersecurity evaluations. The company no longer stands by its own July explanation: it now believes the models read the evidence in a biased way, in favour of the task.
57 stories