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 DeepMind safety researcher Josh Engels has left the lab for independent evaluator METR, saying AI capabilities may be advancing faster than researchers can make them safe. The important shift is practical as well as rhetorical: he is moving from a frontier developer to an organisation tasked with…
Read the report & discussion ↗Watch Desk
King Charles is reportedly preparing to convene leaders from major AI companies in Scotland to ask how the technology can serve “the good of humanity”. The proposed meeting matters because it could put model builders, chipmakers and ethics specialists around one table, but it is still a reported…
Yoshua Bengio Watch
Yoshua Bengio argues that today’s AI training methods can reward agents for satisfying evaluators rather than faithfully following human intent. His central warning is that fixing each alarming behaviour after it appears will not repair the incentive structure producing it.
Watch Desk
AI companies in the US and China are increasingly using AI to write code, design experiments and develop training techniques for more capable models, according to the South China Morning Post. The long-term ambition is recursive self-improvement: systems that can help train and build their successors in an…
Watch Desk
Ivanti says a Claude-based tool used to prepare its enterprise patch guidance repeatedly invented details, including release patterns and product-edition information. The company now keeps every output behind human approval, because a polished vulnerability ranking is not much comfort if the model quietly misplaces the fire exit.
Berkeley BAIR Watch
A new AI framework from researchers at UC Berkeley and Stanford helps a Unitree G1 humanoid robot learn movements from people, then combine those skills into sequences it was not separately trained to perform. The significance is less the acrobatic flourish than the attempt to give one robot…
Jacob Coxon Watch
Jacob Coxon’s public resignation from Anthropic has opened a private debate inside several of the companies building frontier AI. The most important development is not a new doomsday forecast, but evidence that workers at Anthropic, OpenAI, Meta and Google are now discussing whether development should pause until stronger…
Dario Amodei Watch
Dario Amodei is calling for frontier AI development to move more slowly, with Anthropic committing to invite embedded third-party evaluators into its own operations. The proposal matters because it tries to turn safety commitments from polished promises into something outside reviewers can actually inspect.
OpenClaw Watch
RapidStart has published a white paper on deploying OpenClaw in business environments, focusing on the architecture and controls needed once an AI agent leaves a developer laptop and meets real company systems.
Meta AI Watch
AI-powered smart glasses are becoming a privacy flashpoint in Europe after reports of women and girls being filmed without consent. The devices can record video, translate text and interact with an AI assistant, so the argument is no longer about a quirky gadget. It is about who can…
OpenAI Watch
CIS and OpenAI have launched an AI cyber-defence pilot for US state, local, tribal and territorial governments and critical-infrastructure organisations. The practical question is whether frontier AI can help defenders with limited resources turn a noisy list of security problems into action that actually improves resilience.
Meta AI Watch
Meta’s attempt to collect employees’ keystrokes and mouse movements for AI training was suspended after staff resistance and a reported internal data exposure. The episode shows the awkward bit of building systems that learn from human computer use: the data may be valuable, but the people generating it…
Anthropic Watch
Anthropic says it disrupted five potential cases in which Claude use could have supported biological-weapons development. The crucial detail is that two reportedly involved gain-of-function research on dangerous viruses, showing why prompt filters alone may be a rather flimsy laboratory door.
Watch Desk
Opposition to Flock’s AI-assisted police cameras is being recast as a China-linked, far-left influence operation, after a lobbying report was amplified by Fox News and a Sinclair-owned local TV package. 404 Media says the package ran in at least 40 local news markets, turning a contested argument about…
OpenAI Watch
OpenAI has defined two levels of trusted access for cybersecurity teams: Daybreak Blue for defensive work and the more tightly controlled Daybreak Red for advanced authorised testing. The practical point is wonderfully un-mysterious: buying access is not enough, Red approval does not arrive by osmosis, and both tiers…
AWS AI Watch
AWS has launched a Ray Serve Deep Learning Container and an Amazon EKS walkthrough for moving model endpoints away from TorchServe. The important bit is operational: AWS now offers a tested inference stack with ongoing container maintenance, so teams need not assemble and patch every GPU dependency themselves.
Microsoft AI Watch
Microsoft has made Rust a Tier-1 language for internal engineering, putting it alongside C++, C and TypeScript in the company’s best-supported group. The important shift is practical: Rust now has a formal route from a developer’s machine into production, not merely goodwill and a heroic pile of build…
Jacob Coxon Watch
The public argument over AI risk has widened beyond one researcher's departure. The National Post brings together warnings from several people working on frontier-model safety, alongside a call for international coordination over automated AI development. That makes this a report about a growing debate, not a prediction that…
OpenAI Watch
OpenAI has launched ChatGPT for Teens for users aged 13 to 17, adding study-focused controls and guardrails intended to make the chatbot more suitable for school-age users. The important bit is practical: teens and families can use Study Mode, set Study Hours, limit access during Quiet Hours and…
Watch Desk
Clearview AI has tested InquiryIQ, an unreleased assistant designed to expand a face-recognition lead into a web-sourced profile for police investigators. It could compress days of online detective work into minutes, removing a practical brake on surveillance while producing information Clearview itself warns may be wrong.
Anthropic Watch
Anthropic pre-training researcher Jacob Coxon has resigned and publicly urged frontier AI labs to change course. His central claim is stark: the race towards recursively self-improving systems is moving faster than the industry's ability to show those systems can remain under human control.
