UK AISI details deep-access security collaboration with Anthropic and OpenAI

The UK AI Security Institute published a blog post describing an ongoing security partnership with Anthropic and OpenAI, run jointly with the US Center for AI Standards and Innovation. AISI said the two companies had granted its adversarial-machine-learning researchers “deep access” — in-depth model access including non-public tooling and safeguard details that goes beyond what either company discloses publicly — so government researchers could search for vulnerabilities and help fix them before they reached production.

AISI framed the arrangement as running in both directions: it gave governments a clearer picture of frontier-model risk than public documentation allowed, while giving the developers an external check on their own security work from a team the institute said had “expertise across security-critical domains.” Both companies published companion posts describing the same collaboration from their own side, indicating it was a coordinated disclosure rather than a unilateral AISI account.

The post did not name specific vulnerabilities found in this round, nor give a timetable for the work, describing it instead as continuing practice rather than a single completed exercise. It followed AISI’s earlier pre-deployment evaluation of OpenAI’s o1 model with the US institute, extending that relationship from one-off testing before a launch to a standing security function embedded with the developers.

The announcement was one of several around the same period — including a joint US CAISI/UK AISI update from OpenAI three days later — in which government AI safety institutes and frontier labs sought to demonstrate that pre-deployment government access, rather than post-hoc regulation, could substitute for statutory oversight that remained largely absent in both the US and UK at the time.