Anthropic previews a standard for agents to run instruments

Anthropic opened a research preview of the Model Hardware Standard, a framework letting AI agents operate lab and manufacturing instruments such as microscopes and robotic arms in tandem, on any device with a programmable interface, Reuters reported. The worked examples run from drug discovery experiments to laser calibration on a quantum computer.

That is a claim about the last step of a loop. Today's science record covers the steps before it, and the two are worth reading together before deciding how much of the loop is closed.

What the design half currently returns

Anthropic had Claude models design 1,320 protein binders against 15 targets, and 354 of them bound in wet lab tests run by Adaptyv Bio and Twist Bioscience: a 26.8% hit rate, with one target returning no binders at all. The report is the company's own, it has not been peer reviewed, and what it measures is binding, not therapeutic effect.

Roughly three quarters of the proposals failing is not a criticism of the method. It is the shape of the work: generation is cheap, and the assay is what costs money and time. On those numbers, an agent that can also run the assay is worth more than one that designs better, because the wet lab is where the throughput limit sits.

Sungkyunkwan University built the same loop in materials, reported in Advanced Materials: a large language model extracts solid-state synthesis conditions from published literature, proposes routes for a target material, and revises them on experimental feedback. Extract, propose, revise. The instrument is still driven by a person.

The standard is the missing hand

Read the three items in order and the case for a device-control standard makes itself, without help from a press release. Models propose at volume, at a hit rate low enough that the experiment rather than the idea sets the pace. Loops that feed results back already exist. The part not yet standardised is the agent's hand on the device, and a framework for any device with a programmable interface is aimed exactly there.

It is worth naming what today's record does not show. No figure published today measures an agent running an instrument end to end. The 26.8% covers designs handed to two commercial labs to test. The preview is a preview.

The same week, a letter about defences

OpenAI published an open letter signed by more than 100 companies, including Anthropic, Google and Microsoft, arguing that countries and organisations have a closing window to strengthen cyber defences before AI models become capable enough to override them, per the BBC. Anthropic is a signatory, and it previewed device control the same week.

The defences that actually arrived today were plumbing rather than policy. JFrog launched a Software Supply Chain Traffic Controller that intercepts package downloads at the network edge and reroutes them through Artifactory rather than blocking them, aimed at AI coding agents and automated tools pulling dependencies straight from public registries, Investing.com reported. The threat it is built against is documented: the Australian Federal Police arrested two Western Australia men, aged 21 and 23, over alleged membership of TeamPCP, the group blamed for embedding malicious code in hundreds of open source tools from late 2025, Krebs on Security reported.

An agent that pulls a poisoned package writes bad software. An agent holding a programmable interface to a robotic arm is a different exposure, and nothing published today measures it.

What would settle it

• A hit rate for binders designed and assayed inside one agent-run loop, against the 26.8% for designs handed to an outside lab.
• Whether the Model Hardware Standard leaves the research preview with an access model attached, or without one.
• Whether the controls that follow look like the JFrog approach, sitting in the network path where the agent already acts, or like the letter, which asks organisations to move first.