OpenAI Plans Security Sprint With Cathedral to Test Its Defenses

OpenAI is preparing an internal security sprint with its teams, new researchers and partners including Cathedral, using its models to test defenses and identify weaknesses.

The Nov. 2 announcement gave no start date, duration or scope for the planned work. Participants are expected to use OpenAI models to probe the company’s protections and surface weaknesses.

In the previous sprint, one-quarter of OpenAI’s product engineers worked on security remediation. Codex helped teams investigate issues, recreate vulnerabilities in isolated settings and prepare fixes for review.

Cathedral has also taken part in an external security test involving Anthropic. Anthropic said Cathedral researchers reached a limited selection of systems used for sales, human resources and IT functions during an external security test described in a prior account of the test.