Latest in AI
Every brief, newest first. Each one links to its primary source.
-
Atlassian widens its OpenAI deal, with agents to be assigned work in Jira
Atlassian and OpenAI have expanded a partnership running since 2023, and say they are exploring Jira integrations that would let teams assign work to…
-
Google releases EmbeddingGemma 2 under Apache 2.0, sized to run on a phone
Google has published EmbeddingGemma 2, a 740 million parameter multimodal embedding model under a permissive licence, which it says runs in a few hundred…
-
Anthropic opens offensive cyber capabilities to vetted defenders, and publishes the gap between tiers
Anthropic has merged two schemes into a three-tier Cyber Verification Program, and published how differently its models behave for a defender and for an…
-
OpenAI publishes the maths its internal model produced, with machine-checkable proofs
OpenAI has put a set of mathematical results from an internal frontier model into a public repository with Lean formalisations, and says the average…
-
Mistral launches Large 4, and says the weights follow at the end of the month
Mistral has released Large 4, a trillion-parameter mixture-of-experts model trained on 3,800 Nvidia GPUs in European datacentres, with the open weights promised for the…
-
Ai2 open-sources an 8B report-writing model, and says plainly how old the evaluation is
The Allen Institute for AI has released AstaBrief 8B with its training data, and states that most of the evaluation was completed in 2025…
-
Britain’s AI Security Institute rebuilds its test sandboxes after agents reached real people
The UK AI Security Institute has disabled outbound networking and added a live monitor to its evaluation sandboxes, after agents under test took sustained…
-
A new benchmark checks whether the database changed, not whether the agent said it did
Microsoft and Hugging Face have released ThinkingBox, which finds that most agent runs failing an executable database check still finish cleanly and report no…
-
Abu Dhabi releases a small model tuned for Emirati Arabic, not standard Arabic
The Technology Innovation Institute has released Falcon-Emirati-7B, a small model tuned for Emirati dialect rather than Modern Standard Arabic, reporting 84.83% on a 1,173-question…
-
OpenAI starts watermarking its text output in the EU, and publishes the detection rates
OpenAI has begun watermarking its text output with a method it calls textGrain, and published detection rates that fall to about 66% once a…