OpenAI Models Test Its Own Security Defenses
The planned sprint, announced without a timeline, uses AI to find weaknesses in the company's protections before attackers can.
Updated on Nov. 3, 2026 in Artificial Intelligence
OpenAI has announced an internal security sprint in which its teams, new researchers, and partners, including the security firm Cathedral, will use OpenAI's models to probe the company's own defenses and surface weaknesses. The announcement, made November 2, gave no start date, duration, or scope. This is an announced plan, not a completed or ongoing result.
Why it matters
Using a company's own AI models to attack its protections reflects a growing approach in which AI systems are turned on security work itself. It follows an earlier sprint in which a quarter of OpenAI's product engineers worked on security remediation, suggesting the effort is a significant internal commitment.
The announcement cites a previous sprint in which one-quarter of OpenAI's product engineers worked on security remediation, with Codex used to investigate issues, recreate vulnerabilities in isolated settings, and prepare fixes for review. The company has not published measured outcomes from the planned sprint, so its effectiveness remains unquantified.
The players
OpenAI
An AI developer whose frontier models, including the Codex coding model, are the core tools being turned on its own defenses in the sprint.
Cathedral
A security firm that is a named partner in the OpenAI sprint and previously took part in an external security test involving Anthropic.
Anthropic
An AI developer that was the subject of the external security test in which Cathedral participated.
The details
OpenAI plans to have participants use its own models to probe the company's protections and surface weaknesses, a practice known as red-teaming, where testers deliberately attack a system to find flaws before attackers do. In the earlier sprint, the Codex coding model helped teams trace issues, reproduce vulnerabilities in isolated test environments so they could be studied without risk, and draft fixes for human review. Cathedral, a security partner, has previously taken part in an external security test involving Anthropic, during which its researchers reached a limited set of systems used for sales, human resources, and IT functions.
Timeline
November 2: OpenAI announced the planned internal security sprint, with no start date or duration given.
The Tech Race
The effort extends the Codex-assisted remediation sprint that OpenAI ran previously, shifting from repairing known weaknesses toward discovering them. External testing by Cathedral against Anthropic's systems shows that red-team exercises of this kind are already happening across competing AI developers.
Users of OpenAI's products are not expected to see changes from the sprint directly, and the announcement gives no timeline for when findings or fixes might reach them. The practice of using AI models to test AI systems' defenses is still in its early public stage and has not yet produced published results.
The takeaway
Watch for OpenAI to disclose the sprint's scope, duration, and findings, which would show whether AI-driven red-teaming surfaces weaknesses that human teams missed. Track whether Cathedral or other security partners publish results from their participation.
Further reading
For background on the systems driving this work, see Artificial Intelligence coverage.
