In 1984, at the first Hackers Conference, Stewart Brand described two forces pulling at information. It wants to be expensive because it is valuable. It also wants to be free because distributing it keeps getting cheaper.

AI has highlighted another tension. Some biological knowledge could be dangerous in the wrong hands. At the same time, AI is making advanced expertise cheaper and easier to access.

Anthropic is trying to figure out where to draw the line. That puts a private company, at the AI frontier, in the unenviable position of deciding which questions its most capable models will answer.

I doubt Anthropic is comfortable with that power. But the release of Claude Fable 5.1 and Mythos 5.1 shows that it has not found a satisfying way out. Mythos’s biology capabilities are reserved for vetted researchers. Fable 5.1 is available to everyone, but simple, general, questions about biology is enough for Anthropic to pause the conversation and offer an older model instead.

For now, there are other frontier models to choose from. But if one company ever develops an AI so capable that rivals cannot catch up, whoever controls it could shape what billions of people can readily learn, which questions receive answers and which ideas can be explored using the best available intelligence.

Biology makes the danger visible, but it would not stop there. We should decide who gets to draw those boundaries before someone gains the power to draw them alone.

Across society, a question is beginning to follow AI into every industry and occupation: can we let it handle this on its own? Can we even remove the human from the loop?

We will ask that question again and again in the coming years. And, increasingly, the answer will be yes.

Which brings us to Berlin-based AI company Vara. It has received CE certification under the EU Medical Device Regulation for an AI that can report selected screening mammograms as normal without a radiologist reading them. Vara says it is the first breast-imaging AI certified anywhere to do so.

"It took us ten years to get here. What made it possible can become a template for autonomous AI across healthcare," co-founder and CEO Jonas Muff writes in a launch essay.

The autonomy is deliberately narrow. The AI clears only examinations it classifies as clearly normal. Every other mammogram is read by at least one radiologist.

In many organized European screening programmes, two radiologists independently read every mammogram. Around 97% turn out to be normal, according to Vara. Scarce specialists therefore spend much of their time confirming that nothing suspicious is there.

The certification rests partly on ATMON, a Vara-built software system that monitors the AI and acts as a fail-safe. That’s the template Muff is referring to. Because AI performance can shift across scanners, workflows or patient populations, ATMON tracks changes in mammography hardware, system health and daily performance wherever Vara is deployed. If those signals move outside defined limits, autonomous reporting stops there and every mammogram returns to a radiologist.

In PRAIM, an observational real-world study of 461,818 women, the adjusted cancer-detection rate was 17.6% higher with AI support than without, with no increase in recalls. The AI’s safety net prompted radiologists to reconsider examinations they had initially judged normal, leading to 204 cancer diagnoses. Human readers also found 20 cancers in examinations the AI had classified as normal.

PRAIM involved AI-supported double reading, not autonomous screening. Vara has not published prospective results from examinations reported without a human.

Don’t expect this to be implemented overnight, though.

“Screening programmes will not adopt autonomous triage in clinical practice tomorrow, and they are not meant to,” Muff writes. Rollout still depends on national guidelines and clinical readiness in each country.

Vara has shown that autonomous medical AI can be certified under Europe’s demanding medical-device rules. Its route - narrow autonomy, prospective evidence, continuous real-world monitoring and an automatic return to human reading when safety signals move outside set limits - gives other healthcare AI companies a concrete model to study as they pursue carefully defined clinical decisions without human review.

Anthropic says AI models will soon “make important contributions to scientific discovery”. It points to early results from Claude Fable 5.1 and Claude Mythos 5.1 as evidence.

They are the same underlying model with different safeguards. Fable is available to everyone. Mythos is the less restricted version for vetted life scientists and cyberdefenders, currently limited to selected US organisations.

Its strongest biology demonstration is protein design. Anthropic gave Mythos open-source design and folding tools, then had two outside organisations test its proposed binders - proteins designed to attach to chosen targets. Across 12 targets, nearly 50% of the designs bound, compared with the 10-15% Anthropic says is typical today.

Anthropic also compared Mythos with results from competitions run by Swiss protein-engineering company Adaptyv Bio. Anyone can submit designs created using any method; Adaptyv selects some, physically makes them and tests them under the same laboratory protocols, creating a public wet-lab benchmark. On three targets, Anthropic says Mythos’s best binders attached around ten times more tightly than the strongest comparable competition designs.

This is a company-reported demonstration, not a peer-reviewed study. Anthropic has not published a technical report on the binder campaign, and binding is only an early step towards a useful drug.

Then, as has sadly become custom by now, there are the safeguards. Anthropic says Fable now flags benign elementary biology and medical questions 85% less often than Fable 5 did at launch. But life-science research queries are still sent to an older Opus model. When I asked whether Fable 5.1 could discuss biology without switching models, it began to answer, then paused the conversation, tagged it “[bio]” and offered to continue with Opus 5.

Biologists are not amused. Stanford geneticist Anshul Kundaje called Fable “essentially unusable” for biology and Anthropic’s strategy “absolutely ridiculous”.

A general AI model using specialist tools to produce experimental binders at this hit rate would be meaningful progress. But the launch makes two cases at once: Anthropic’s models are becoming useful for science, while its safeguards keep their strongest biological capabilities away from most scientists.

Reading a lung-cancer tissue sample gives pathologists a long to-do list. Is there enough tumor for genetic testing? Which type of lung cancer is it? Does it show protein markers that could affect treatment, and at what levels?

A new system called LUCAID turns that list into a team effort between AIs. A team led by Charité and Aignostics uses a large language model as foreman: it sends each job to a specialist AI module, then pulls the answers into a report. The work is described in a preprint.

The researchers tested LUCAID on tissue samples from 70 people with lung cancer. LUCAID and five experienced pathologists assessed the cases independently. The doctors later reviewed them together and agreed on what the answers should be. Across 328 individual decisions, LUCAID matched the group’s answers 93% of the time. Each doctor’s original answers matched them between 68.3% and 81.1% of the time.

While the results sound impressive, this was a small early study, and parts of the setup may have favored LUCAID. Larger, independent studies are needed.

LUCAID shows how a general AI might supervise a bench of narrow tools and become something closer to a digital colleague.

Drug-discovery company Om has proclaimed that “vibe discovery is coming.” To that end, it has built a connector - Om MCP - that lets Codex and Claude Code use its platform. Once added, you can ask the agent to research a drug target, run models that predict protein structures or how molecules might bind, and bring the results back into the same session. Om bills it as a way to run multi-step workflows without stitching together APIs yourself. You need an Om account, and some actions are paid.

☑️ The naked emperors of AI drug discovery.

☑️ Insilico releases five specialist LLMs for drug discovery.

☑️ AI-designed fluorescent protein tags work in mammalian cells and at 75°C.

☑️ BigHat doses first patient with a drug from its AI platform.

☑️ ARPA-H commits up to $98.5 million to rare-disease AI.

☑️ Arc’s State virtual-cell model is published in Cell.

☑️ Boehringer licenses Owkin’s K Pro AI scientist and biomedical data.

☑️ C-PLANK learns from molecule-protein interactions inside living cells.