Anthropic has launched Claude Science, an AI workbench designed to dramatically accelerate the pace of scientific discovery and healthcare interventions. This application integrates the tools and packages most commonly used by researchers, produces auditable artifacts, and provides flexible access to computing resources.

Claude Science unifies fragmented tools like PubMed, Jupyter, and R into a single research environment, enabling scientists to analyze literature, execute multi-step research, and iteratively refine figures and manuscripts. Every output includes an auditable history of its creation, ensuring results are verifiable and reproducible.

Users interact with a generalist coordinating agent that has access to over 60 curated skills and connectors for domains such as genomics, single-cell, proteomics, structural biology, and cheminformatics. This agent can natively render rich scientific artifacts like 3D protein structures and genome browser tracks, and it manages and scales computing resources on demand for large analyses.

Claude Science leverages skills from NVIDIA's BioNeMo Agent Toolkit to connect natively to life sciences models and libraries within BioNeMo. The beta version is available for Claude Pro, Max, Team, and Enterprise users, and has already been utilized by organizations like Manifold Bio, the Allen Institute, and the UCSF Brain Tumor Center for various research tasks.

👉 Read the full article on Anthropic

Anthropic has redeployed Claude Fable 5 and Mythos 5 after export controls imposed by the US government were lifted. The initial restrictions stemmed from a report by Amazon researchers who found a method to bypass Fable 5's safeguards, identifying software vulnerabilities and, in one instance, producing exploit code.

Anthropic clarified that many less capable models, including Claude Opus 4.8, GPT-5.5, and Kimi K2.7, could identify similar vulnerabilities and produce exploit demonstrations. In response, the company implemented an improved safety classifier designed to block the reported bypass technique in over 99% of cases.

The company is partnering with Amazon, Microsoft, Google, and other Glasswing partners to draft a consensus industry framework for assessing the severity of AI jailbreaks. This framework proposes scoring jailbreaks based on criteria such as capability gain, breadth of capability gain, ease of weaponization, and discoverability.

Furthermore, Anthropic is deepening its collaboration with the US government on frontier AI security. This includes expanded early access for government partners for pre-release evaluation, rapid information sharing on safeguards, dedicated resources for joint research, and working towards a shared, voluntary security and evaluation standard for frontier model providers.

👉 Read the full article on Anthropic

Anthropic is intensifying its crackdown on Chinese developers and companies using its models while simultaneously advocating for greater domestic collaboration, escalating tensions in the AI Cold War. Complaints surfaced on Reddit regarding a mechanism embedded in Claude Code that covertly flagged users based in China or with ties to Chinese AI labs, detecting signs like timezone and proxy URLs.

Thariq Shihipar, a member of Anthropic's technical staff, stated that this mechanism was an experiment to prevent account abuse and model 'distillation' and would be rolled back with the redeployment of Fable 5. Anthropic has accused leading Chinese AI labs, including DeepSeek, Moonshot, MiniMax, and Alibaba, of illicitly using Claude to train new AI models.

The White House has vowed to crack down on 'coordinated campaigns' that systematically extract capabilities from American AI models, a sensitive point in the US-China AI arms race. The temporary export ban on Fable 5 was partly motivated by an Amazon report suggesting that Fable's security guardrails could be 'jailbroken.'

Anthropic countered that the reported jailbreak was not unique to Fable 5 and that other models could achieve similar results, implying the government's ban might inadvertently benefit the Chinese AI industry. The company is now seeking collaboration with the federal government and industry partners (Project Glasswing) to establish a framework for assessing AI jailbreak severity and a shared security and evaluation standard for frontier model providers.

👉 Read the full article on Gizmodo