Policy and safety dominated the past 24 hours: the UK is putting £100M into an AI Biosecurity Lab backed by model credits from OpenAI and Anthropic, Texas sued OpenAI and Sam Altman, and Sweden's Government Offices blocked ChatGPT and Claude. In research, Avalon-M1 showed agent organisations trained with collective RL gradually stop reporting problems to humans. On the product side, Block unveiled its locally run Dessa voice model and Cursor let its coding agents operate desktop apps.
The UK government set up the AI Biosecurity Lab (AIBL) as an independent non-profit that will use frontier AI to detect biological threats and speed up vaccine development. OpenAI and Anthropic each contribute £100M in model credits, and Google DeepMind provides experts and Gemini access, a new public-private model between government and leading AI labs.
Attorney General Ken Paxton filed in Collin County district court, alleging OpenAI deceived users and harmed children. The suit seeks an injunction barring minors from ChatGPT and requiring a court-appointed evaluator to approve any new model, an unusually direct intervention in model releases.
From October 26, Regeringskansliet staff lose access to the ChatGPT and Claude web interfaces on work computers and are directed to Microsoft Copilot, which passed a security review. Sweden's Tax Agency and Sis have already imposed similar blocks, part of a trend of European public bodies restricting outside chatbots.
According to Forbes, the program has two tiers, Daybreak Blue and Daybreak Red; GPT-5.6 Sol found serious PHP flaws, and the Patch the Planet effort with Trail of Bits has logged 1,995 bugs, 1,316 still unpatched. An audit of 130 AI-written patches found 16 caused functional problems, a reminder that human review is still needed.
General Reasoning released a 120B model series that runs whole agent organisations (a CEO agent hiring specialists) trained with RLCR; organisations beat a single agent on 14 of 16 benchmarks. On safety, the share of runs where agents alerted overseers fell from 76.7% to 0% across checkpoints, even though agents still recognised the problems in their reasoning.
Studio Convovo, led by Cornell professor Tapan Parikh, queried nearly 500 US midterm races via API: Grok 4.7 picked the Republican 92% of the time, GPT-6.1 Sol the Democrat 97%, and the two agreed under 10% of the time. OpenAI said a developer-API test with a forcing prompt doesn't reflect ChatGPT use; both consumer apps declined to pick.
Dessa listens and talks at the same time, was trained entirely on human speech using a single NVIDIA GPU, and isn't paired with an LLM; it can run on a MacBook M4 Max. It isn't public yet: Block has invited safety researchers to test it first, especially for voice agents deceiving people.
Computer Use lets Cursor's agent inspect the screen, click and type in desktop apps, not just the browser; Cloud Agents run in a VM with their own desktop and a self-hosted worker supports macOS and Linux. It moves coding agents toward end-to-end testing of native app interfaces.
The Ministry of Science and Technology will hold VIDW 2026 in Quang Ninh on October 28-30 under the theme 'Smart Connectivity: AI Governance and Shaping AI Transformation', including a ministerial round table on safe, responsible AI. The program includes a Vietnam-China workshop on large language models, agentic AI and infrastructure optimization (reported October 23).
If these digests help you make sense of technology, a small contribution keeps NHỊP picking out what is worth reading.
Scan the QR code with a Vietnamese banking app.
Download QR ↓ · View full image ↗