AI-found bugs are expected to rise, and OpenAI has a five-stage response plan
OpenAI has proposed a five-stage plan for dealing with serious security flaws in widely used software. The plan came from a two-month effort to respond to flaws in PHP, an open-source programming language used to run websites and web applications. Calif found one of the PHP flaws using OpenAI’s GPT-5.6 Sol model through OpenAI’s Patch the Planet project, which aims to improve the security of open-source software. Wiz also found both flaws independently and reported them.
“These bugs are serious, but they are among several other serious security flaws we have already found and are working to fix. We expect to find many more similar flaws in the coming months,” OpenAI wrote.
The plan is designed for security flaws that affect many products at the same time. Software distributors and service providers need time to prepare fixes before details become public. But that time could become very short if widely available AI tools, including tools attackers can use, get close to the abilities of the most advanced AI models that find the flaws.
Five stages, with software maintainers in charge
In the first stage, software maintainers, security researchers, and technical partners confirm that a flaw exists, work out how serious it is, and build and test a fix. As part of testing, they use the most advanced AI models to look for ways around the fix and to find similar flaws.
In the second stage, organizations affected by the flaw get extra time to prepare. A software distributor might adapt the fix for an older version. A hosting provider might test a way to reduce the risk. A security company might create web application firewall rules, which filter out harmful traffic before it reaches an application.
In the third stage, the people responding to the flaw bring in government agencies. During the PHP response, OpenAI and its partners worked with the White House’s Super Intelligence Force, the Office of the National Cyber Director, CISA, and international partners. In the fourth stage, once the fix and plans for sharing information are settled, maintainers notify more software distributors, service providers, and system operators. In the fifth stage, after the fix is released, the response team publishes advice and continues helping organizations install it.
Who gets information first, and what they receive
OpenAI’s plan uses three factors to decide which organizations get information early: the role they play in protecting systems, how much time they need to take action, and whether they can safely handle sensitive information. The same rules would apply to similar organizations, and software maintainers make the final decision. Each participant receives the information needed for its job. This could range from network logs showing signs of an attack to working examples that demonstrate how to exploit the flaw.
“Even limited details can reveal a vulnerability,” OpenAI wrote. Teams using the plan would review the timeline and make changes if attackers begin using the flaw, information about it leaks, or a public software update reveals too much. They would also reconsider the schedule if delays in fixing the flaw increase the risk to users.
What OpenAI wants help with
OpenAI is asking software maintainers, researchers, and system operators for advice on how long organizations should have to prepare before details of a flaw are made public, how to balance early warnings to many organizations against the risk of leaks, and how to fix older software packages that no longer receive security updates. OpenAI also raises a more difficult question: should governments or system operators use a vulnerability to fix systems they own or are allowed to manage if they have lost administrative access? OpenAI describes a possible future in which systems across the internet are scanned continuously, with the process either warning website owners about problems or fixing the systems automatically.