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
A bipartisan Senate permitting bill would prevent utility ratepayers from bearing the costs of new power and grid upgrades needed for AI data centres, Axios reports. The proposal also reaches far beyond AI, changing how infrastructure projects and environmental reviews are handled.
Read the report & discussion ↗NVIDIA Watch
NVIDIA has announced a reference design for securing autonomous AI agents, pairing a sandboxed runtime with a separate hardware watchdog, according to Moor Insights & Strategy. The design shifts some safety controls out of the model itself and into the systems that run it, a practical change for…
Meta AI Watch
Meta’s Muse can take on tasks such as buying groceries and negotiating bills, but a BBC report argues that AI agents doing this at scale could strain the websites and services they use. The question is no longer only what an agent can do for its owner, but…
Watch Desk
The Trump administration has unveiled America.gov, an AI-powered portal intended to help people find federal information and services through ordinary-language questions. Its initial role is navigation, not completing applications: a useful distinction when the destination is a passport or Medicare service, not a restaurant recommendation.
Google DeepMind Watch
Google DeepMind has announced SynthID Bio, a method for embedding detectable markers in AI-designed proteins. The potential payoff is a way to identify some AI-designed biological sequences even after they have been physically made, rather than relying on digital records alone.
Yann LeCun Watch
Yann LeCun says he has “zero concerns” about AI causing human extinction, arguing instead that recent agent incidents point to poor oversight and badly designed systems. In a Fortune interview, the AI researcher also warns that safety arguments could be used to entrench the biggest companies, while describing…
Anthropic Watch
Anthropic has made Claude for Government generally available to US federal and state agencies in a FedRAMP High-authorised environment. The practical draw is a government-focused setup with spending caps, department-level administration and audit controls, rather than simply a familiar chatbot in a new suit.
Watch Desk
Researchers and others concerned about artificial intelligence gathered at Lighthaven in Berkeley for AGI.WTF, a conference on AI risk and the people building a movement around it. The story is not just about worst-case forecasts: it offers a glimpse of how that community is organising, debating and trying…
Watch Desk
Building an AI agent that can use real services means solving more than login: teams must manage whose account it acts on, what it may access and how risky actions are approved. A Firecrawl guide compares nine authentication platforms and separates the job into four distinct problems, a…
Google DeepMind Watch
Google is preparing Gemini 4 for launch amid internal disagreement about how well it handles practical coding tasks, Bloomberg reports. The gap at the heart of the story is a familiar one in AI: strong benchmark results do not automatically translate into a model that works well on…
Watch Desk
Developers told VentureBeat that safeguards in OpenAI and Anthropic models are flagging ordinary aerospace, robotics and security tasks, interrupting work and sometimes pushing them towards other models. The dispute is practical: how can AI tools block dangerous requests without treating routine engineering as a threat?
Microsoft AI Watch
Microsoft 365 Copilot Business customers will move to usage-based billing by default for eligible services from 2 November, while everyday AI tasks remain covered by the per-user subscription. The practical catch: administrators must first create a spending policy before any usage-based charges can begin, and then need to…
Google DeepMind Watch
Google DeepMind has announced Gemini 4 Argon, a frontier model designed for lengthy professional tasks, with a one-million-token output limit and an initial focus on trusted cyber defenders. The launch pairs ambitious performance claims with a phased rollout, an introductory API price and striking examples of AI doing…
Geoffrey Hinton Watch
Geoffrey Hinton estimates there is a 10% chance AI could wipe out humanity within the next decade, according to Futura’s account of an interview with BBC Newsnight. He also stresses that the risk is impossible to assess accurately, which makes the figure a stark personal estimate, not a…
OpenClaw Watch
OpenClaw’s 2026.9.7 release combines performance improvements with safer upgrades and new sign-in options. The project says it is snappier under load, smoother in long chats, and adds beta support for the OpenAI Agents API and ChatGPT sign-in.
Microsoft AI Watch
Microsoft says healthcare organisations are putting AI into everyday workflows, with examples ranging from drafting clinical notes to helping staff search research and business information. The useful detail is that these are specific deployments, not just another promise that healthcare will be transformed shortly after lunch.
MIT CSAIL Watch
An AI system developed by researchers from MIT, Carnegie Mellon, NYU and Stanford has beaten top Stratego players, including the world’s strongest player, while using far fewer training examples than its predecessor. The result is a striking advance in hidden-information game-playing, with possible lessons for other strategic decisions…
AWS AI Watch
AWS has published a working example of three AI agents collaborating on a music-production pipeline using Amazon Bedrock AgentCore Runtime Instances. The practical difference from its serverless MicroVM option is substantial: Runtime Instances can host multiple agents together, provide GPU access and preserve a workflow across sessions lasting…
