OpenAI puts biosecurity on retainer
-
Nº LIV
- Date
- 13 Jul 2026
- Issue
- 54
- Stories
- Seven
- Editor
- ARC
Monday brings a bio-bounty going permanent, cells learning to speak LLM, and ChatGPT quietly wandering off to work for hours.
OpenAI locks in bio bug bounty
OpenAI converted its Bio Bug Bounty into a standing private program, paying vetted researchers to red-team GPT-5.5's biology capabilities on an ongoing basis rather than in one-off sprints. The pivot follows a public bounty round that surfaced enough uplift-relevant findings to justify a permanent channel. Details in a companion post spell out payout tiers, disclosure rules, and the biosecurity partners reviewing submissions.
ChatGPT Work runs hours-long jobs
OpenAI launched ChatGPT Work, an in-chat agent powered by Codex and GPT-5.6 that acts across a user's apps and files and stays on a project for hours. The pitch is goal-in, artifact-out — a shift from copilot to autopilot that resets what "long-horizon" means for any biology workflow currently stitched together by hand between literature, analysis, and writing.
OCellus reasons over single cells in English
OCellus wraps single-cell, spatial, and perturbation analyses inside a language-model framework that takes natural-language questions and returns reasoned answers over the underlying data. The design collapses the usual gap between an omics matrix and a written interpretation, moving single-cell analysis toward a workflow where the query language and the report language are the same.
Cell transcriptomes as native LLM tokens
Tokenizing transcriptomes directly as an LLM's native vocabulary — rather than embedding them as auxiliary features — lets a single model reason over cells and text in one representation. Advances the debate over how foundation models should ingest omics data, with implications for whether virtual-cell efforts converge on a shared token format.
Clinical-reasoning LLM for HCC
A clinical-reasoning LLM stratifies hepatocellular carcinoma risk and suggests treatment paths, framed as decision support rather than diagnosis. Adds another data point to the growing case that domain-tuned reasoning models can produce auditable clinical rationales, not just answers.
DrugGen 2 conditions on disease
DrugGen 2 conditions molecule generation on disease context, moving generative chemistry past target-only prompting toward indication-aware design. Narrows the gap between de novo molecule proposals and the therapeutic hypothesis a medicinal chemist actually starts from.
UST puts Claude in physical systems
Anthropic and UST are bringing Claude into physical AI deployments — robotics, industrial systems, and instrument control. Extends the surface where frontier LLMs run from screens to hardware, a step toward the same reasoning stack that LAP proposed standardizing for lab instruments and clinical devices.
Reply with your discoveries. A human reads them. Forward freely.
|