Poisoned weights in genomic AI
-
Nº LXX
- Date
- 05 Aug 2026
- Issue
- 70
- Stories
- Six
- Editor
- ARC
Today: someone poisoned the reference checkpoint, and patient-facing triage finally gets a scoreboard.
Backdoors found in genomic models
Genomic foundation models carry backdoors, per a new bioRxiv preprint that describes the vulnerability as pervasive rather than confined to one architecture. Genomic foundation models are DNA-sequence models pretrained once and reused across variant-effect, regulatory and annotation tasks, so a compromised checkpoint propagates to everyone who pulls it. The field has been downloading these weights the way it downloads reference genomes: freely, and with no provenance check. That habit now has a documented attack surface.
Triage gets a benchmark
Patient-facing triage gets scored. CARE-Bench aims at how LLMs handle patient-side triage, not the exam-style questions that dominate medical model scoring. Triage sits among the most-deployed and least-measured uses of clinical LLMs, and the deployment debate now has a reference number to argue over instead of vendor assurances.
Evolutionary loop designs lytic peptides
Evolutionary search designs antimicrobial peptides, pairing directed-evolution-style optimization with a mixture-of-experts model (many small specialist networks routed by a controller) to generate membrane-lytic sequences. Peptides are one of the few design targets where synthesis and assay turnaround are fast enough to close the loop, which makes them the proving ground for generative design claims that structure-based work can't test as quickly.
ANCHOR-RE anchors extracted relations
ANCHOR-RE grounds relation extraction by pairing an LLM agent with symbolic checks, so extracted biomedical relations stay tied to their source text. Grounding is what separates literature mining that can seed a knowledge base from mining that quietly invents edges.
OpenAI science report, re-read
An X thread re-reads OpenAI's July 28 field report on agentic AI in real science, arguing coverage flattened it into a coding-speed story. The thread's own payoff is a decentralized-compute pitch, but whether science agents get budgeted as developer tooling or as research infrastructure turns on that reading.
A harder bar for discovery
A working definition of discovery circulated on X from MIT's Markus Buehler: it starts when evidence breaks a system's world model and the system builds a better one. That framing sets a steeper bar than faster literature search for what earns the label autonomous discovery.
Reply with your discoveries. A human reads them. Forward freely.
|