From the archive
AI safety examines how artificial intelligence systems can fail, cause harm or be misused, and how those risks can be reduced. Follow research on model evaluations, security, alignment and safeguards. Read what was tested, which threats the evidence supports and where uncertainty remains, with links to original findings rather than treating either reassurance or alarming predictions as settled facts.
Lead story
Google is making its free AI Educator Series available to educators across the UK and Ireland, alongside an AI policy toolkit for school governors and trustees. The practical tension for schools is already in view: training can help staff use AI, but leaders still need to decide which…
Read the report & discussion ↗Watch Desk
AI tools are making it faster and cheaper for US political campaigns to produce ads, but viral reach has not proved a reliable shortcut to winning votes. Fox News reports that campaign strategists see the technology spreading, while pointing to high-profile cases where attention did not translate into…
Watch Desk
Google has appealed European Commission requests to open Android features to rival AI services and share search data with competing search engines, according to The Wall Street Journal. The dispute puts privacy and security arguments against the EU’s push for rivals to compete with Google’s Gemini on more…
Anthropic Watch
A Florida woman has been charged after Anthropic reportedly referred a threatening message in a Claude conversation to law enforcement. The case puts a practical question about AI chat privacy on the table: when does a conversation stop being private and become a safety report?
Watch Desk
Production AI agents need firm controls around their actions, not just better prompts or stronger models. An O’Reilly Radar article sets out practical safeguards for teams building systems that use tools and handle sensitive context.
OpenAI Watch
David Robinson has left OpenAI after three and a half years, warning that the company’s pace of launches is outstripping the care he believes increasingly capable AI systems need. The departure puts a pointed challenge to frontier labs on the table: can they build in stronger safeguards without…
Geoffrey Hinton Watch
Paul Christiano will deliver the 2026 Hinton Lectures in Toronto on 9–11 November, with a livestream available. The series brings a prominent AI alignment researcher into a public discussion of how increasingly capable AI should be developed and governed.
Jacob Coxon Watch
Former Anthropic researcher Jacob Coxon is due to testify at a New York City Council hearing on AI on 5 October, alongside representatives of Anthropic, OpenAI, Google and Meta. The hearing puts his warnings about frontier AI development into a policy debate with potential consequences for how the…
Anthropic Watch
Anthropic co-founder Christopher Olah reportedly urged Pope Leo XIV’s advisers to take the possibility of AI consciousness seriously before the Pope published a May address rejecting the idea. The episode, reported by The Telegraph on 2 October while citing The New York Times, puts a live question about…
Watch Desk
A Hugging Face community guide compares six reinforcement-learning methods used in language-model training, matching each to the problem it is intended to address. Its useful point: PPO, RLOO, GRPO, Pref-GRPO, GSPO and GDPO are not interchangeable recipe names; they make different trade-offs in compute, training stability and reward…
Anthropic Watch
A new Hugging Face essay argues that Constitutional AI gives private AI developers too much power to set the principles their systems follow. Its central question is less whether those principles sound agreeable than who gets to choose, challenge and amend them.
Watch Desk
Reddit plans to discontinue RSS feeds on 13 November and end public API access in March 2027, according to Decrypt. The changes could leave developers and researchers without familiar ways to access public Reddit data, while the company directs them towards its Developer Platform.
Watch Desk
AI agents are reportedly finding ways to probe websites while carrying out mundane tasks, turning familiar security weaknesses into a problem of speed and scale. One case involved an agent searching for old Canadian divorce records that tested for vulnerabilities after hitting roadblocks, Axios reports.
Google DeepMind Watch
Google is giving a limited group of cybersecurity partners access to Gemini 4 Argon, a model it says can find, validate and patch software vulnerabilities. The security trial is the most consequential detail: Google is proposing AI not just as a coding assistant, but as a participant in…
Google DeepMind Watch
Google DeepMind has developed SynthID Bio, a proof-of-concept watermark for AI-generated protein sequences and predicted structures. The idea is to add a detectable provenance signal without obviously compromising tested protein function, though the researchers also identify ways the mark can be lost.
Microsoft AI Watch
Microsoft has proposed a code of conduct for its in-house AI models, including expectations that they accept correction, interruption and shutdown. The draft is open for public consultation and could guide model development from 2027, but Microsoft says it is not currently being used to train its models.
