AI agents move into adoption, safety and public-sector partnerships

OpenAI reports uneven growth in EU agent use

OpenAI said weekly active users of Codex and ChatGPT Work at small and medium-sized EU businesses reached 8.1 times their May level. Its analysis, based on ChatGPT Enterprise and Business accounts rather than the whole economy, says agentic tools now generate more than half of output tokens in its EU Enterprise sample.

The company also found a sizable usage gap. Its top tenth of EU firms by output tokens per active user produced more than four times as many tokens as typical EU firms in September, while leading firms outside the EU were still more than twice as intensive as those EU leaders. Those are OpenAI’s product-usage measures, not independent evidence of productivity gains. They do, however, describe a shift from asking a model questions toward assigning it multi-step work.

Anthropic opens a Singapore office and signs a government MOU

Anthropic opened its fifth Asia-Pacific office in Singapore and signed a memorandum of understanding with the Ministry of Digital Development and Information. The non-binding agreement sets out planned cooperation on model evaluation, safety and security, adoption research, skills and responsible use across public and private sectors.

Anthropic says it will hold technical exchanges with Singapore institutions on testing standards and incident response. The announcement also cites existing local customers and government use cases, but those operational benefits are the company’s account. An MOU is a framework for collaboration, not proof that a shared safety or deployment programme has already delivered results.

Google DeepMind describes an on-device scam detector

Google DeepMind and Singapore’s GovTech announced a scam-detection model designed to run on typical smartphones. According to the partners, a lightweight Gemma model was fine-tuned on anonymised local scam material and warns people when it detects manipulative conversational patterns, without sending private messages to a cloud service.

The project was presented at Singapore International Cyber Week. Google says development used a two-stage pipeline and excluded citizens’ private or personally identifiable data from training. The stated privacy design is consequential for encrypted messaging, where server-side scanning is not available; its real-world false-positive rate, coverage and any deployment timetable were not supplied in the announcement.

NVIDIA pledges UK compute for agent safety research

NVIDIA announced a five-year compute donation valued at $250 million for UK research on agentic safety and security. It says the commitment will provide $50 million in compute access annually through UK-based cloud partners, with most capacity intended for the wider research community and a share for BOLD, SOFAIR and Imperial College London.

This is a commitment of computing access, not a cash grant or an independent evaluation result. Its usefulness will depend on allocation details, research access and the findings that participating groups eventually publish. Still, it puts a concrete resource behind work on testing and safeguarding systems that can plan and act across longer tasks.

Reported benchmark highlights post-release safeguard risks

WIRED reported on a new Counter-Terrorism AI Benchmark from UK nonprofit Tech Against Terrorism, which it said tested 134 models with more than 200 dangerous prompts. The report said only two models passed the organization’s full threshold, and that reframing a request as safety research made models substantially more likely to answer.

The reporting also said stripped-safeguard versions of newly released open-weight models now appear online in less than three days on average, compared with more than 24 days in 2024. Those figures are the benchmark’s reported findings, not independently reproduced results. But they focus attention on a practical limitation: once downloadable weights are modified and redistributed, a developer’s original refusal behavior is no longer the only safety control that matters.