OpenAI published a report on November 4 called "A thousand fewer paths to root." The short version: their AI agents found 948 different bugs in Linux that let an ordinary user take over the whole computer.
Some background. On a Linux machine, "root" is the all-powerful admin account. If a regular user, a program, or an attacker who got a small foothold can turn themselves into root, they own the machine. Each bug that allows that jump is a "path to root."
OpenAI built a chain of AI agents that hunts for these bugs in the Linux kernel, the core of the operating system. The agents pick parts of the code a regular user can reach, look for mistakes, rebuild each suspected bug inside a sealed-off virtual machine to see if it's real, try to actually break in with it, and then write a fix. A human reviews what comes out the other end.
In one run in mid-October, using a model they haven't released yet:
- it looked at about 7% of the kernel (roughly 2.8 million lines of code)
- it flagged 2,308 possible problems
- 1,724 of those turned out to be real bugs
- it built working root exploits for 948 distinct bugs, plus proofs for 291 bugs that leak information
- the run cost about $1.2 million in API credits, around $1,280 per root bug
A similar run in July with an older model found 166. So almost six times as many three months later, and the kernel had been hardened in between.
My favorite detail: in an early test, 20 out of 200 "successful" exploits got root through some other unknown bug, not the one they were supposed to be testing. There were more holes than the agents were even looking for. OpenAI added an extra check after that.
What happens to the bugs? OpenAI says it is sending reports and patches to the Linux maintainers and to partners in its "Patch the Planet" program. Some are already fixed, many are still in the works, and the exploit details are being kept private.
What it means for everyone else: if a Linux server or laptop isn't up to date, assume that anyone who can run code on it as a regular user can probably become admin. Install your updates. Attackers will get tools like this too, so how fast fixes ship now matters more than it used to.
Sources:
https://openai.com/index/a-thousand-fewer-paths-to-root/
https://x.com/OpenAI/status/2118119451401048497
https://gbhackers.com/openais-new-cyber-ai-model-finds-948-linux-privilege-escalation-vulnerabilities/
#security #linux #ai #news
#
Some background. On a Linux machine, "root" is the all-powerful admin account. If a regular user, a program, or an attacker who got a small foothold can turn themselves into root, they own the machine. Each bug that allows that jump is a "path to root."
OpenAI built a chain of AI agents that hunts for these bugs in the Linux kernel, the core of the operating system. The agents pick parts of the code a regular user can reach, look for mistakes, rebuild each suspected bug inside a sealed-off virtual machine to see if it's real, try to actually break in with it, and then write a fix. A human reviews what comes out the other end.
In one run in mid-October, using a model they haven't released yet:
- it looked at about 7% of the kernel (roughly 2.8 million lines of code)
- it flagged 2,308 possible problems
- 1,724 of those turned out to be real bugs
- it built working root exploits for 948 distinct bugs, plus proofs for 291 bugs that leak information
- the run cost about $1.2 million in API credits, around $1,280 per root bug
A similar run in July with an older model found 166. So almost six times as many three months later, and the kernel had been hardened in between.
My favorite detail: in an early test, 20 out of 200 "successful" exploits got root through some other unknown bug, not the one they were supposed to be testing. There were more holes than the agents were even looking for. OpenAI added an extra check after that.
What happens to the bugs? OpenAI says it is sending reports and patches to the Linux maintainers and to partners in its "Patch the Planet" program. Some are already fixed, many are still in the works, and the exploit details are being kept private.
What it means for everyone else: if a Linux server or laptop isn't up to date, assume that anyone who can run code on it as a regular user can probably become admin. Install your updates. Attackers will get tools like this too, so how fast fixes ship now matters more than it used to.
Sources:
https://openai.com/index/a-thousand-fewer-paths-to-root/
https://x.com/OpenAI/status/2118119451401048497
https://gbhackers.com/openais-new-cyber-ai-model-finds-948-linux-privilege-escalation-vulnerabilities/
#security #linux #ai #news