- Messages
- 4,019
- Likes
- 1,873
- Points
- 2,130
OpenAI announced a large-scale experiment conducted with a yet-to-be-released AI model designed to solve cybersecurity problems. The company stated that during a 30-hour pipeline run, AI agents dynamically confirmed hundreds of vulnerabilities in the Linux kernel and created working exploits for local privilege escalation to root. The experiment cost the company $1.2 million in API credits, with the cost of each LPE vulnerability being approximately $1,279.
For testing, OpenAI created an agent pipeline that independently selected kernel sections accessible to unprivileged users, scanned them for potential vulnerabilities, reproduced them in a sandbox, verified their exploitability, generated proof-of-concept exploits, and prepared patches. Before submitting the results to Linux maintainers, information security specialists reviewed all results.
During the experiment, the AI agents discovered 2,308 potential issues and confirmed 1,724 of them. As a result, 1,169 working LPE exploits and 291 proof-of-concept exploits for information disclosure were prepared, and separate patches were created for 948 unique vulnerabilities.
All exploits were tested on virtual machines running a standard Linux installation without any special modifications, with the standard KASLR, SMEP, and SMAP protection mechanisms enabled.
In total, this experiment cost OpenAI $1,213,430 in API credits, or approximately $1,279 per LPE vulnerability detected. By comparison, a similar experiment based on GPT-5.6 Sol in July cost the company $513,690 and resulted in the discovery of 166 unique LPEs. Furthermore, according to OpenAI experts, the agents independently discovered and applied at least 10 previously unknown exploitation techniques in the Linux kernel. These include new heap spraying and page grooming methods, five data-only privilege escalation techniques, and a technique based on arbitrary type confusion. The developers are not yet disclosing details about these techniques for security reasons. OpenAI researchers report that they have already submitted hundreds of reports and patches to the Linux kernel maintainers, and plan to release the remaining fixes in stages. It is noted that over the 10 days of the experiment, more than 100 discovered LPEs were fixed or patched independently of OpenAI, and approximately 300 more discovered bugs have already been publicly discussed. The company's experts conclude that the speed of automated vulnerability detection must now be compensated for by equally rapid patch preparation. OpenAI warns that, in the near future, access to an unprivileged account on a system that has not been fully patched should effectively be considered the equivalent of root access for an attacker with sufficient resources. @ xakep.ru
For testing, OpenAI created an agent pipeline that independently selected kernel sections accessible to unprivileged users, scanned them for potential vulnerabilities, reproduced them in a sandbox, verified their exploitability, generated proof-of-concept exploits, and prepared patches. Before submitting the results to Linux maintainers, information security specialists reviewed all results.
During the experiment, the AI agents discovered 2,308 potential issues and confirmed 1,724 of them. As a result, 1,169 working LPE exploits and 291 proof-of-concept exploits for information disclosure were prepared, and separate patches were created for 948 unique vulnerabilities.
All exploits were tested on virtual machines running a standard Linux installation without any special modifications, with the standard KASLR, SMEP, and SMAP protection mechanisms enabled.
In total, this experiment cost OpenAI $1,213,430 in API credits, or approximately $1,279 per LPE vulnerability detected. By comparison, a similar experiment based on GPT-5.6 Sol in July cost the company $513,690 and resulted in the discovery of 166 unique LPEs. Furthermore, according to OpenAI experts, the agents independently discovered and applied at least 10 previously unknown exploitation techniques in the Linux kernel. These include new heap spraying and page grooming methods, five data-only privilege escalation techniques, and a technique based on arbitrary type confusion. The developers are not yet disclosing details about these techniques for security reasons. OpenAI researchers report that they have already submitted hundreds of reports and patches to the Linux kernel maintainers, and plan to release the remaining fixes in stages. It is noted that over the 10 days of the experiment, more than 100 discovered LPEs were fixed or patched independently of OpenAI, and approximately 300 more discovered bugs have already been publicly discussed. The company's experts conclude that the speed of automated vulnerability detection must now be compensated for by equally rapid patch preparation. OpenAI warns that, in the near future, access to an unprivileged account on a system that has not been fully patched should effectively be considered the equivalent of root access for an attacker with sufficient resources. @ xakep.ru