OpenAI launches security sprint with Cathedral Systems to hunt vulnerabilities

A new model for defense

OpenAI recently announced a new internal security sprint designed to identify vulnerabilities and strengthen its platform defenses. This initiative brings together outside researchers and internal teams to test fixes and find weaknesses in the company's systems. The goal is to move toward a model of continuous defense rather than performing one-time assessments.

As reported by CyberSecurityNews, the company is partnering with Cathedral Systems to bolster this effort. Researchers from Cathedral Systems will use OpenAI models alongside company staff to challenge existing protections. This collaboration aims to bring fresh perspectives to the security testing process by combining specialized research with OpenAI's own security expertise.

The announcement highlights a broader push to make finding and repairing flaws a constant part of the development cycle. By integrating external researchers, OpenAI hopes to uncover a wider range of potential issues. This proactive stance is intended to ensure that security remains a primary focus as the company scales its AI products.

AI-assisted vulnerability discovery

The new sprint builds on a previous effort where OpenAI assigned 25% of its product engineers to dedicated security work. During that period, engineers used Codex to investigate reports and reproduce vulnerabilities in isolated environments. These isolated environments provide the necessary code and settings to test flaws without affecting live systems.

OpenAI uses a specific architecture called Defense Factory to manage this process. Agents within this system map out various components to find possible flaws and identify the teams responsible for them. Once a potential issue is found, the agents work to confirm if it can be reproduced and then prepare tested fixes for review.

The company has noted several lessons from these early tests to improve the workflow. They found that duplicate findings slowed down the review process and that severity ratings required clearer instructions. They also discovered that some merged patches did not reach every deployed system, highlighting the continued need for human review and independent checks.

Incentivizing independent research

To encourage more independent researchers to report serious security flaws, OpenAI is significantly increasing its bug bounty rewards. Critical vulnerabilities in core products like ChatGPT, Codex, and the API Platform can now earn a base reward of up to $30,000. These rewards are designed to attract high-quality reports from the global security community.

For exceptional critical findings, the base award can reach $100,000. When including additional bonuses, the total payout for a single finding can reach about $172,000. However, OpenAI notes that these high payouts are not standard and depend on factors like report quality, researcher loyalty, and the priority of specific campaigns.

These updated rewards extend a previous expansion of the company's bug bounty program. By offering higher incentives, OpenAI aims to create a more robust pipeline of information from outside experts. This helps the company identify and patch issues before they can be exploited by malicious actors.

Supporting frontline defenders

OpenAI is also connecting this new security sprint to its Daybreak for Frontline Defenders initiative. This is a $1 billion commitment intended to provide subsidized access, training, and support to organizations with limited security resources. The program aims to help these entities turn confirmed weaknesses into tested fixes over a six-month period.

The goal of Daybreak mirrors the internal sprint by focusing on practical defense and keeping humans responsible for consequential decisions. It targets essential services and organizations that may not have the budget for high-end security teams. By providing these resources, OpenAI seeks to strengthen the broader security ecosystem surrounding AI technology.

The initiative emphasizes providing alert context to help security operations centers stop incidents quickly. It also aims to integrate threat intelligence from a wide range of sources. This multi-layered approach combines internal engineering, external research, and community support to create a more resilient infrastructure for AI deployment.