Labour peer Baroness Twycross has said the UK's AI Security Institute identified dozens of vulnerabilities that developers fixed before their models were publicly released.
In a written parliamentary answer on 9 October, Twycross said the institute had secured privileged access through trusted, voluntary relationships with leading artificial intelligence laboratories.
She said these relationships allowed the institute to keep pace with rapidly changing technology and adapt its work as models developed.
Twycross added that the institute was in daily contact with industry partners and worked with international allies on risks posed by increasingly capable AI systems.
The answer responded to a question about access to frontier models for pre-release safety testing. It did not identify the individual vulnerabilities, name the affected models or announce a new compulsory testing requirement.
