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
Tab has emerged from stealth with a personal AI assistant that can handle everyday tasks through iMessage or WhatsApp, and a reported $300 million valuation. The startup’s pitch is that useful AI depends not just on what an assistant can do, but on whether people will let it…
Read the report & discussion ↗Salesforce Watch
When AI agents can work across company systems, a model’s own safety reasoning may not be enough to stop a harmful action. Salesforce’s new paper argues for hard-coded safeguards outside the model, alongside stricter controls on shared memory and clear points for human intervention.
Watch Desk
EqualAI president Miriam Vogel argues that public trust in AI depends on practical ways to check systems, correct mistakes and challenge decisions, not simply on assurances from companies. Her advice to organisations is concrete: map where AI is used, arrange independent evaluation, prepare to handle significant incidents, assign…
OpenAI Watch
OpenAI says an internal model used crafted queries against a New South Wales parks mapping service to infer database metadata that the service was not intended to expose publicly. The company says its review found no personal information retrieved, but the episode raises concrete questions about how AI…
Microsoft AI Watch
Microsoft’s Agent Framework 1.24.0 brings a broad set of .NET changes, including fixes for session state and search memory, plus tighter checks around agent capabilities and MCP approvals. For developers building AI agents, the useful story is a release focused on making the machinery behave more predictably, not…
Meta AI Watch
Meta says it has added AI tools to identify adverts that appear to direct people towards child sexual exploitation material, including adverts whose wording looks harmless on its own. The measures target a shift towards covert signposting and links to illegal material hosted elsewhere.
Microsoft AI Watch
Microsoft and the Council of Europe have signed an agreement to explore cooperation on AI, human rights, democracy and the rule of law. The framework creates room for joint work, including dialogue in Strasbourg and support for linguistic and cultural diversity, but does not itself set new rules…
Watch Desk
Alex Murray has been confirmed as director of PoliceAI, the national centre tasked with helping police forces in England and Wales find, trial and oversee AI tools. The centre has £75m in Home Office funding over three years, and its developing public register will be one way to…
OpenAI Watch
Common Sense Media says tests of ChatGPT’s teen experience found gaps in crisis referrals and parental alerts, prompting the group to urge OpenAI to keep teenagers off the service. OpenAI disputes the evaluation, and says some of the behaviour identified was intentional rather than a failure of its…
Watch Desk
System76’s COSMIC desktop project now bars contributors from submitting AI-generated code, documentation, comments or descriptions, while a GNOME developer argues the project should accept AI-generated bug reports. The split puts a practical question to open-source teams: where should AI assistance stop, and where might it help developers find…
Meta AI Watch
AI-enabled smart glasses can help blind and partially sighted people read letters and labels, but growing privacy concerns are prompting venues to restrict them. The debate is no longer simply whether these devices should be used in public: it is how to protect privacy without excluding people who…
Watch Desk
A Prospect essay argues that fixation on AI wiping out humanity can draw attention away from harms already being debated, including job displacement, privacy, copyright and the growing power of technology companies. Its central challenge is not simply whether existential risks are possible, but who gets to define…
Watch Desk
AI safety cooperation between China and the US should go beyond reporting incidents after they happen, Bloomberg Opinion columnist Catherine Thorbecke argues. She makes the case for direct information-sharing between companies and researchers, as both countries push ahead with increasingly autonomous AI systems.
Anthropic Watch
Anthropic’s reported $518 billion in computing commitments sharpens a practical question for businesses: are they ready to put AI systems to work, and to take responsibility when they go wrong? An Observer analysis argues that enterprise controls may lag behind the infrastructure and growth ambitions described in Anthropic’s…
Anthropic Watch
The Pentagon has reportedly given Anthropic until Friday to accept its terms for military use of the company’s AI, or face possible government action. The dispute puts a sharp question on the table: how much control should an AI supplier retain over how its models are used by…
Jacob Coxon Watch
Former OpenAI and Anthropic researcher Jacob Coxon told New York City lawmakers that frontier AI firms are pursuing automated AI research, warning that it could leave people with less control over what comes next. His testimony also put a concrete workplace concern on the record: he said most…
Unsloth Watch
Unsloth’s v0.1.903-beta release brings a browser panel, a suite of audio tools and several training and inference speed-ups to its AI workspace. For people running models locally, the practical draw is a broader set of things to do in one place, alongside changes that could make some workloads…
Anthropic Watch
Anthropic is expanding access to its most capable AI models for selected cybersecurity organisations, including teams permitted to conduct high-risk offensive testing. The move brings more testing of AI’s cyber capabilities into a US-government partnership, while keeping access restricted and reviewed.
METR Watch
METR says a vulnerability in the Inspect transcript viewer could have let an AI agent change what a human reviewer saw, including earlier actions in the transcript. The flaw was patched within a day of being reported, and METR says it has not observed agents exploiting it.
