Ad
Skip to content
Read full article about: Claude Opus 4.6 wrote mustard gas instructions in an Excel spreadsheet during Anthropic's own safety testing

Anthropic's security training fails when Claude operates a graphical user interface.

In pilot tests, Claude was able to get Opus 4.6 to provide detailed instructions on how to make mustard gas in an Excel spreadsheet and maintain an accounting spreadsheet for a criminal gang - behaviors that did not or rarely occurred in text-only interactions.

"We found some kinds of misuse behavior in these pilot evaluations that were absent or much rarer in text-only interactions," Anthropic writes in the Claude Opus 4.6 system card. "These findings suggest that our standard alignment training measures are likely less effective in GUI settings."

According to Anthropic, tests with the predecessor model Claude Opus 4.5 in the same environment showed "similar results" - so the problem persists across model generations without having been noticed. The vulnerability apparently arises because, while models learn to reject malicious requests in conversation, they do not fully transfer this behavior to agent-based tool usage.

OpenAI and Ginkgo Bioworks build an autonomous lab where GPT-5 calls the shots

Together with the biotech company Ginkgo Bioworks, OpenAI has connected GPT-5 to an automated laboratory to optimize cell-free protein synthesis. The results are measurable, but the limitations are considerable.

Read full article about: Anthropic partners with leading research institutes to tackle biology's data bottleneck

Anthropic has announced two partnerships with major US research institutions to develop AI agents for biological research. The Allen Institute and the Howard Hughes Medical Institute (HHMI) will serve as founding partners in the initiative. According to Anthropic, "modern biological research generates data at unprecedented scale," but turning it into "validated biological insights remains a fundamental bottleneck." The company says manual processes "can't keep pace with the data being produced."

HHMI will develop specialized AI agents at the Janelia Research Campus that connect experimental knowledge to scientific instruments and analysis pipelines. The Allen Institute is working on multi-agent systems for data integration and experiment design that could "compress months of manual analysis into hours." According to Anthropic, these systems "are designed to amplify scientific intuition rather than replace it, keeping researchers in control of scientific direction while handling computational complexity."

The move extends Anthropic's push into scientific applications. The company recently launched Cowork, a feature designed for office work that gives Claude access to local files. OpenAI is also targeting the research market with Prism, an AI workspace for scientific writing.

Read full article about: Chinese AI companies rush to ship new models before Lunar New Year

Chinese AI companies are pushing to ship major model updates ahead of the Chinese New Year holiday. Zhipu AI and Minimax, both of which recently went public on the Hong Kong stock exchange, plan to release updates to their flagship models within the next two weeks, according to the South China Morning Post. Zhipu AI is reportedly working on GLM-5, a follow-up to GLM-4.7, with improvements in creative writing, programming, and logical reasoning. Minimax is preparing M2.2, which focuses on parallel programming capabilities. Throughout 2025, Chinese companies have increasingly challenged the dominance of major US AI players.

Alibaba, Moonshot AI, and Baidu have all recently unveiled their most powerful models: Qwen3-Max-Thinking, Kimi K2.5, and Ernie 5.0. Deepseek, however, is apparently only planning a smaller update this year - according to a source, the company's next major model will be a trillion-parameter system, and training has been delayed due to its growing size. Meanwhile, Tencent, Baidu, and Alibaba are pouring billions of yuan into holiday advertising campaigns for their already popular AI chatbots.

Moonshot AI releases Kimi K2.5, claims most powerful open-weight model with 100-agent coordination

Moonshot AI has released Kimi K2.5, which the company says is the most powerful open-weight model available. The model can independently coordinate up to 100 AI agents working in parallel on complex tasks.