Research
Biological and chemical research, manufacture and agriculture depend on a set of computational services whose degradation would hamper ordinary work and would slow any emergency response. Some of these services, among them the catalogs on which the supply of key reagents depends, at present have very little security. We expect that they will become vulnerable to attack by capable open-weight models during late 2026 or early 2027.
We have identified and are continuing to prioritize the elements of this infrastructure that matter most. Working with the security staff of a participating institution, and using the most capable model to which we have access together with conventional security tooling and human review, we inventory the institution's systems, their connections and their access controls, examine the relevant code and configurations, and rank candidate vulnerabilities by severity. We reproduce the serious ones in isolated test environments, prepare the code or configuration changes that repair them, review and install those changes with the institution's own staff, and then repeat both the failure tests and the ordinary operating tests.
We do not name the institutions on this list, and we do not publish the list.
The services concerned are few, are identifiable in advance, and are at present largely undefended. Repairing them before capable attacks arrive is cheaper than repairing them afterwards, and it keeps working the infrastructure on which both ordinary research and any emergency response depend. An institution whose security staff would be willing to take part in this work is invited to write to us.