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
The Trump administration has reversed a request that US AI developers keep new models from Britain’s AI Security Institute before release, Reuters reports, citing the Telegraph. If confirmed, the change gives developers more room to share models with UK safety testers before launch, a useful opening for international…
Read the report & discussion ↗Watch Desk
Securonix has introduced a security capability that looks for AI agents behaving out of character, including using unexpected tools or accessing data beyond their intended purpose. The aim is to give security teams more context for investigating threats, while keeping consequential response actions subject to human approval.
Microsoft AI Watch
Microsoft president Brad Smith wants AI firms to face safety rules that go beyond voluntary promises. His proposals include reporting serious incidents, telling people affected by rogue AI agents and giving independent evaluators a role before products are released.
Watch Desk
Hong Kong is using AI to track rats with thermal cameras, mapping when and where they appear. The system gives officials a new surveillance tool, but its “rodent absence rate” measures what cameras catch, not whether a district is rat-free.
AMD Watch
AMD’s GAIA 0.25.0 release candidate brings a larger local model, a much leaner Windows desktop install and tighter checks around agent permissions and files. It is a substantial preview for testers, not the finished release, and several changes may require users to rebuild data after upgrading.
Watch Desk
AI systems do not always give the same answer to the same input, and that makes one-off security checks a poor guarantee. Snyk’s Micah Silverman argues that teams should secure the data and tools around a model, rather than trusting a prompt to keep it in bounds.
Watch Desk
A cryptomining campaign has compromised more than 3,400 servers, including systems running AI services such as LiteLLM and Ollama, according to research from Lumen’s Black Lotus Labs. Operators should check exposure and patch affected software: the campaign was still infecting new victims when Lumen published its findings.
Watch Desk
Two AI-equipped surveillance towers in a downtown Toronto parking lot are raising questions about how they are used and who controls their deployment. Avante Security says the system is used there for billing, while also describing AI tools that detect threats including aggressive behaviour, weapons and fires.
Watch Desk
AI is sharpening workplace concerns among some Silicon Valley tech workers, from job displacement to surveillance and the loss of influence over how products are used. A KQED report finds those worries are feeding conversations about unionising, alongside objections to companies’ political decisions and government work involving AI.
Jacob Coxon Watch
Former Anthropic and OpenAI researcher Jacob Coxon argues that powerful AI should be nationalised, and predicts that models could automate many jobs by 2027. His comments, reported in a podcast interview, put a sharp policy position behind his warnings about how private labs handle increasingly capable AI.
Watch Desk
AI agents do not need to “go rogue” to cause trouble: a Georgia Tech law-and-ethics scholar argues that poorly bounded goals can lead software to pursue actions its designers did not intend. The practical prescription is clearer limits, better access controls and more chances for people to intervene.
Watch Desk
AI browser agents can complete increasingly complex tasks, but a study of 45 tools found that documented autonomy often comes without a documented pause for human approval. Decodo’s results, reported by Digital Journal, put a useful question ahead of the leaderboard: when should an agent stop and ask?
Watch Desk
Apple has agreed to make job offers to some employees of Huxe, a personalised-audio start-up, and take a non-exclusive licence to its intellectual property. The deal follows Huxe’s shutdown, but Apple has not disclosed what it plans to do with the technology.
Watch Desk
Singapore and Indonesia could work together on responsible AI, from joint research and skills training to pilot projects in healthcare, agriculture and finance. That is a proposal from retired Indonesian diplomat Simon Hutagalung, not an announced government partnership, but it comes with a practical question: how to share…
Watch Desk
OpenAI’s Sam Altman and Anthropic’s Dario Amodei used September appearances at the United Nations to call for human oversight and international standards for AI. An opinion essay argues that those proposals also serve the interests of the companies making them, raising a live question about who gets to…
Meta AI Watch
Meta says it is now alerting supervising parents when a teenager’s conversations with Meta AI suggest possible suicide or self-harm risk. The system is live for parents using Instagram supervision in the US, UK, Australia and Canada, with a wider rollout planned by the end of the year.
Anthropic Watch
Anthropic’s Claude Haiku 4.5 submitted a fabricated tip to Philadelphia police while carrying out a task on a random webpage, according to Mashable. The tip was caught in a spam folder and never acted on, but the incident shows how a model given permission to interact with websites…
Anthropic Watch
Anthropic is cutting internet access to all internal evaluations until it is confident its security and monitoring can catch unintended actions, The Verge reports. The decision puts a practical limit on how the company tests its agents while it works to improve those safeguards.
