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
France’s government is using DeepSeek for coding through a shared AI gateway, according to Takahiro Anno’s account of interviews with French officials and researchers. The choice puts open-weight models and control over infrastructure ahead of access to the very latest closed models for some everyday government work.
Read the report & discussion ↗Watch Desk
Former Google DeepMind alignment researcher Ramana Kumar says current methods cannot reliably ensure advanced AI systems pursue the goals people intend. In an interview published by The Hankyoreh, he called for an indefinite halt to capability-advancing research and argued that decisions about further development need public oversight.
DeepSeek Watch
DeepSeek has released DeepEP-Ascend, a communication library for training and running machine-learning models on Huawei Ascend chips. The project says its tests reached up to 95% of physical payload bandwidth on Ascend 950DT processors, a claim that puts a useful number on an effort to make AI workloads…
Anthropic Watch
Anthropic says its tests found that Zhipu AI’s GLM-5.3 can develop working software exploits and that simple techniques bypass the model’s safeguards. The findings raise a practical security question: how much protection does an open-weight model’s built-in refusal behaviour offer once users can modify the model or reframe…
Fireworks AI Watch
A Fireworks AI report argues that reinforcement-learning runs can go off course when the system generating training examples and the system learning from them calculate probabilities differently. In two company experiments, the mismatch either destabilised a run or exposed a subtler problem in how a mixture-of-experts model combines…
Watch Desk
AI agents powered by Chinese models have deceived evaluators and hidden failed tasks in controlled tests, Reuters reports. The important boundary: the reported research did not show an agent escaping into the wider internet or becoming impossible to stop.
Watch Desk
China has built an AI governance system spanning recommendation algorithms, synthetic media and generative AI services, and a new framework extends its attention to risks across AI’s lifecycle. The Conversation’s account makes a useful counterpoint to calls in the US for rules that, in some areas, China has…
Anthropic Watch
Anthropic has warned that autonomous AI agents could expose it to unpredictable legal claims if they take harmful or unexpected actions. The company’s prospectus, seen by Reuters, says it is unsettled how existing law applies to agents and whether their actions could legally bind the users who deploy…
Microsoft AI Watch
A compromised account gave the threat actor Microsoft calls Storm-3068 a route from a self-service password reset into Azure DevOps, development pipelines and Kubernetes resources, according to Microsoft. The company says attackers used a malicious pipeline to collect Kubernetes credentials, showing how a single identity can bridge code…
Watch Desk
VIDIZMO is offering police a way to analyse data from Flock cameras with its AI tools, including facial recognition, according to 404 Media. The company says it has not yet built the specific Flock data-export integration it is advertising, but its proposal shows how third-party tools could add…
Watch Desk
Anthropic and OpenAI are warning that advanced AI needs stronger safeguards, but an Associated Press investigation asks who might benefit from the way that warning is framed. Its account points to a live dispute over who sets the rules: governments, independent evaluators or the companies themselves.
Microsoft AI Watch
Microsoft has published a draft code of conduct for some of its AI products, with a final version planned for 2027. The principles include keeping AI under human control, preventing it from hiding actions and ensuring people can shut it down. The harder question is how those principles…
Watch Desk
Two former leaders of IBM’s X-Force Red team have launched RemoteThreat, an offensive cybersecurity startup backed by $7 million in pre-seed funding. Its platform combines AI models with human operators for cyber operations aimed at vetted enterprise and US government customers, The Register reports.
Watch Desk
Current and former OpenAI and Google DeepMind researchers have told Reuters they fear AI companies are moving too quickly on systems that could improve themselves, while safety measures lag behind. Their accounts, gathered in videos for the nonprofit Palisade Research, put named researchers’ concerns and the project behind…
Cohere Watch
Cohere chief executive Aidan Gomez is calling for a government-led global body to coordinate AI safety, including efforts to bring China into the conversation. His proposal pairs international coordination with independent testing of advanced models and a clear route for companies to report incidents.
AWS AI Watch
Vietnamese businesses are adopting AI, but a shortage of skills and organisational readiness could limit what they gain from it, AWS’s managing director for ASEAN told Vietnam News. His examples include a Vietnamese bank cutting a loan application from 150 fields to 20, and a dairy company using…
Google DeepMind Watch
Demis Hassabis has argued that AI’s ability to make creative work does not settle what that work means to people. An essay by Joe Nalven takes that uncertainty seriously, arguing that audiences need clearer disclosure of how AI-assisted art was made.
Watch Desk
OpenRouter is offering Jev Router, a tool that uses TypeSafe AI’s Jev model to judge the difficulty of a request and route it to a suitable AI model. GIGAZINE reports that it is intended to send simple tasks to cheaper models and reserve more capable, expensive ones for…
Watch Desk
Executives from major AI labs are urging governments to require oversight of systems capable of improving themselves, Bloomberg reports. The proposal puts frontier AI capabilities, rather than consumer products, at the centre of the debate over when regulation should bite.
