The life sciences verification programme gives approved teams more permissive classifiers. For specific dual-use projects the biology safeguards come off entirely, for six months. The price is 30 days of data retention for monitoring.
- Standard Use is for the whole team and is renewed once a year.
- High-risk Use is for one project, for six months; for Mythos, only a small circle for now.
- The cyber classifiers stay in place under both types of access.
A virologist making a vaccine and someone who wants a virus to spread more easily ask almost the same questions. The classifier often cannot tell them apart by the words, so until now it stopped both.
Anthropic changes the question. Instead of “what are you asking” it checks “who are you”.
What actually changes
Until now the safeguard stood at every door and questioned every request. Now it lets the verified in and reviews the record afterwards, to catch a pattern spread across many sessions.
The risks are named by them: a hijacked account, an employee working for someone else, and an agent that does something dangerous during a long task without anyone asking it to. The third is the newest. Vetting the person does not vet the agent they leave running overnight.
And one small but important detail - the 30 days of retention. For a lab that is a small price. For a company guarding someone else's intellectual property, it is a line in the contract that some lawyer will want to read twice.
If you work in a lab that has been getting refusals for normal work, the path is through the waitlist. If you do not, the news for you is different: the boundary is no longer only in the model but also in the vetting of the people around it.