AI's biology risk stops being hypothetical
-
Nº XCVIII
- Date
- 14 Sep 2026
- Issue
- 98
- Stories
- Six
- Editor
- ARC
Monday: a threat report, a protein tuned past evolution, and three models arguing about what drugs do to cells.
Anthropic details how people tried to misuse Claude
Anthropic published its most detailed threat intelligence report yet, cataloguing how people have tried to misuse Claude for cyberattacks, influence operations, surveillance, biology, and weapons development. Biology appears as a named misuse category, which is itself the news: a frontier model vendor is now publishing what it observes people attempting on the bio side rather than asserting that safeguards work. The disclosure lands amid wider alarm about AI-designed viruses and the thin machinery available to check any of it, with Axios reporting this week that there is still no reliable way to know how these systems will behave. Documented attempts give the biosecurity guardrail debate something concrete to argue over instead of scenarios.
Reinforcement learning optimizes a protein for brightness
Reinforcement learning pushes a protein past what evolution ever optimized for. CreiLOV is a light-sensing protein, and brightness was never the trait selection scored it on, so a protein language model trained on natural sequences has no reason to favor bright variants. Adding reinforcement learning (training against a reward for a chosen property, instead of copying existing examples) reorients the model toward the trait a designer wants, the thread argues. That moves protein language models from imitating evolutionary statistics to optimizing objectives nature never ran, which is where most therapeutic and reagent design actually lives.
GEM-GPT designs therapies cell type by cell type
GEM-GPT resolves therapeutic design down to individual cell types, a bioRxiv preprint reports, pointing generative drug design at systems pharmacology rather than single targets. Interventions are proposed against a patient-level picture of which cell types are doing what. Single-cell atlases made that resolution visible years ago, while drug design has mostly kept predicting at tissue average. Closing that gap on the design side is the claim worth watching here.
ShEPhERD-2 designs molecules from interaction patterns
ShEPhERD-2 generates molecules from interaction profiles, the pattern of contacts a ligand makes with its target, rather than from scaffolds or fingerprints. Using contacts as a shared representation lets generative design carry across targets instead of retraining for each one.
scDEFT predicts drug effects and tests counterfactuals
scDEFT predicts drug effects and answers counterfactuals: what a cell would have done under a different treatment. Adding that second question moves single-cell drug modeling from response prediction toward the causal comparisons trial design actually rests on.
General AI forecasters get tested on glucose data
Time-series foundation models get tested on continuous glucose monitoring, with dietary logs added as context. Benchmarking general pretrained forecasters (trained broadly, not built for this task) against a physiological signal sets a reference point for whether generic models earn a place in metabolic monitoring.
Reply with your discoveries. A human reads them. Forward freely.
|