Expert-enhanced pathogens

This risk does not run through the amateur but through those who already know. The question is whether biological design models allow a competent team to produce an agent with properties that nature does not offer —and which therefore has no countermeasures ready.

What has been demonstrated is the preceding step. King and co-authors published in Science, on 6 August 2026, sixteen viable bacteriophages designed with genomic language modelsLLM (large language model)A large language model: the kind of AI behind assistants such as ChatGPT, Claude or Gemini, trained on enormous amounts of text to predict which word comes next.For exampleLike your phone's autocomplete, but trained on vastly more text: that is why it can carry a whole conversation and not just the next word. on the ΦX174 template ^[597]Generative design of bacteriophages with genome language modelsKing, Samuel H.; Hie, Brian · 2026 · paperView source ↗verified through CrossrefAccessed on 9 September 2026. And Anthropic reports that Mythos 5.1 outperforms a pre-trained protein language model at predicting AAV capsid assembly using only its reasoning ^[82]System Card: Claude Fable 5.1 & Claude Mythos 5.1Anthropic · 2026 · system cardView source ↗Accessed on 9 September 2026. Genome design works, and general reasoning is starting to catch up with specialised tools.

What is missing to reach the risk is everything else, and the labs say so. Anthropic places Mythos 5.1 at CB-1 —significant help in synthesising a known weapon— and explicitly below CB-2, the threshold of replacing the scarce expert, which is the limiting factor for novel weapons ^[82]System Card: Claude Fable 5.1 & Claude Mythos 5.1Anthropic · 2026 · system cardView source ↗Accessed on 9 September 2026. On OpenAI’s Critical axis, zero of three evaluations cross the threshold ^[825]GPT-5.6 System CardOpenAI · 2026 · system cardView source ↗Accessed on 9 September 2026.

What this does not demonstrate. The phage work was built so as not to produce a human pathogen: it started from non-pathogenic systems in laboratory strains of E. coli, and the Arc Institute states that Evo cannot generate human viral sequences because of deliberate exclusions from the training data ^[93]How We Built the First AI-Generated GenomesArc Institute · 2025 · institutional blogView source ↗Accessed on 9 September 2026. The success rate also constrains the reading: Simon Jackson calculates that around 5% of the designs worked, and Jordi García Ojalvo points out that designed genomes have to be tested one by one ^[976]Expert reaction to generative design of bacteriophages with genome language modelsScience Media Centre · 2026 · pressView source ↗Accessed on 9 September 2026. The CB-2 threshold is not a measurement but a boundary defined looking forward.