OpenAI and Claude in the Biology Lab: The AI industry warns about viruses

Runaway chatbots, AI-generated threats to police officers, manipulated bank accounts. New examples of what can go wrong are added almost daily as increasingly powerful AI systems encounter the real world. These warnings no longer come only from politicians and AI critics. The developers themselves also paint darker scenarios. In mid-September, Anthropic substantiated these fears with its own security report. The company based in San Francisco revealed that it had blocked targeted misuse attempts by state and politically motivated actors. Donald Trump after a meeting with tech CEOs in the White House. The superintelligence agreement is legally non-binding – but "morally binding" according to Trump. © imago When the chatbot becomes a lab assistant Five cases from biological research were particularly alarming. In one, a user wanted to use the AI model Claude to specifically alter the Chikungunya virus. In another, the goal was to better adapt avian influenza to mammals. Anthropic reportedly blocked the involved accounts and informed state authorities. "Biological misuse is one of the most serious risks of groundbreaking AI models," the company warns in the report. Trump's "morally binding" AI pact: How the tech elite should self-regulate Bio-labs in Ukraine: The thin line between research and bioweapons Behind this lies the so-called dual-use dilemma. Knowledge and methods that serve the development of vaccines and medicines can also be misused for biological weapons. AI could facilitate access to this knowledge – even for people without the appropriate training. A laboratory and practical experience do not replace it yet. The fear of biological misuse was already voiced by industry giants like Sam Altman and Dario Amodei in an open letter to the US Congress in June. They demanded stricter legal controls over the ordering of synthetic DNA and RNA in the future. Who is warning, researches with The AI corporations are now pushing this tension even further. In April alone, it became known that Anthropic had acquired the biotech startup Coefficient Bio for around $400 million in shares, according to media reports. In the same month, the head of the pharmaceutical giant Novartis was appointed to its board of directors. Furthermore, the company sought experts for proteins and nucleic acids through job advertisements. In mid-September, the official presentation followed. Anthropic presented its own physical biology lab in the San Francisco Bay Area. There, Claude supports biological research on-site. The AI searches large DNA datasets, develops hypotheses, and suggests experiments that scientists then conduct in the lab. The company presented a first research result immediately. Around 950 parallel Claude agents searched genetic data within 21 hours. One of them discovered an unusual DNA pattern, which ultimately led to the identification of a previously unknown enzyme system with CRISPR-like features after further analysis and laboratory tests. Google DeepMind and its sister company Isomorphic Labs are also advancing AI-supported biology. In early October, their involvement with Meta in the Virtual Biology Initiative, founded by Mark Zuckerberg and Priscilla Chan, was announced, which was previously announced in April. The expanded initiative, involving several US agencies, includes commitments worth $1.8 billion – including money, data, computing power, and measurement technology. An employee is pipetting cell culture medium at the Center for Medical Biotechnology at the University of Duisburg-Essen. AI is increasingly expected to support biological research. © imago The so-called Virtual Biology Initiative is understood as international, cross-sector cooperation. Its goal: to generate vast amounts of biological data and develop AI models from it that can predict cell behavior – such as how they react to drugs or genetic interventions. "This investment in generating biological data will help create an open, standardized data commons," promises Pushmeet Kohli, Vice President for AI for Science at Google DeepMind. Security based on trust Parallel to this entry into physical research, the tech elite is working on the political framework for their AI models. At the end of September, after confidential meetings and exclusive dinner rounds in the orbit of the tech elite, US President Donald Trump finally presented the so-called "White House Superintelligence Agreement." A paper without legal force, without penalties for violations, legally non-binding, but as Trump himself emphasized, "morally binding." The document provides that corporations should conduct internal security controls, establish oversight by their own leadership bodies, and involve external auditors. However, the selection and appointment of these auditors lie with the companies themselves. No state control body or independent certification authority is provided for in the agreement. The audit reports do not have to be published either. "From aviation to banking, high-risk sectors are subject to independent oversight and sensitive penalties," the AI researcher Heidy Khlaaf from the AI Now Institute criticized in September in the journal Nature. "AI companies should not be an exception." The State as Partner A state body still exists in the US: the Center for AI Standards and Innovation, or Caisi. However, it does not consider itself a regulatory authority but rather an interface between the government and the AI industry. Together with the companies, the agency tests models and develops "voluntary security standards" – also for biosafety risks. Caisi should also prevent foreign governments from basing American technologies on "onerous and unnecessary regulation." OpenAI's Sam Altman is on the UN Security Council, and Anthropic's Dario Amodei is also involved. Both demand stricter legal controls over the purchase of synthetic DNA and RNA. © Kyle Mazza/imago On the other side of the Atlantic, the European Union is much more determined. With the AI Act, it is relying on legally binding rules. "Europe has the power to act. We decide our rules – not Big Tech," stated Commission President Ursula von der Leyen in September regarding Europe's digital policy. But the major providers are also at the table in the concrete design. The Code of Practice sets out how the requirements of the law for particularly powerful general-purpose models should be implemented. It regulates, for example, how providers should document training data, respect copyrights, and assess and limit security risks for particularly powerful models. Scientists and civil society, as well as Anthropic, Google, Microsoft, and OpenAI were involved in its development. When security determines the market As necessary biosafety regulations can become a competitive factor, the researchers Jacob Beal and Tessa Alexanian already showed in 2025 with the example of synthetic DNA and RNA providers. Their analysis was made on the occasion of the US regulation on "improving safety in biological research." As leading AI developers later demanded, they call for binding minimum standards for controlling this genetic material. At the same time, the industry should contribute its expertise to the development of the rules, independent bodies to check and certify their compliance – but ultimately only the state can enforce this. Anthropic CEO Dario Amodei: "The pace of AI progress must be controlled" Trump's "morally binding" AI pact: How the tech elite should self-regulate Even then, the researchers found purely voluntary requirements insufficient. Without an agreed standard, governments hesitated to enforce screening, fearing that it would stifle innovations, favor regulatory influence, or fragment international cooperation. But binding rules do not only solve a security problem. They also change the competition. Providers must invest resources in their control systems and can fall behind competitors who scrutinize less strictly. Conversely, the early introduction of new standards can later become an advantage: once binding rules are in place, pioneers will have an incentive to support their enforcement because it "represents a competitive advantage for them and a disadvantage for their competitors." Your opinion is important to us. Send us your feedback on this article to: briefe@berliner-zeitung.de