OpenAI used AI to find 948 ways to hijack Linux computers

The short version

• OpenAI pointed a team of AI agents at the core of Linux and found 948 flaws that let an ordinary user take over a whole machine.
• The run cost about $1.2 million, and the AI also wrote draft fixes that it is sending to the people who maintain Linux.
• OpenAI's advice is blunt: if an unpatched system lets anyone log in, even with a basic account, treat it as fully exposed.

What OpenAI did

Linux runs most of the world's servers and cloud computers. It also sits under every Android phone. Its core has more than 40 million lines of code.

OpenAI built a chain of AI agents, which are programs that can carry out steps on their own. Some agents read the code and picked the parts worth checking. Others hunted for flaws in those parts. A third group tried to trigger each flaw on a test machine, to prove it was real.

In mid-October, one run produced 2,308 suspected flaws. The AI confirmed 1,724 of them on a test machine. It then built working attacks for 948 flaws. Each one lets a user with a basic account become the all-powerful "root" user. That means full control of the computer.

The test machines ran a normal desktop setup, with standard protections switched on. They were not weakened to make the job easier.

The AI also drafted a fix for each flaw. Human security researchers reviewed the results before anything went to the Linux maintainers. OpenAI says it has already sent hundreds of reports and fixes for the most critical flaws.

What is actually new

The scale and the price are new. In June, OpenAI reported 24 such attacks found with an earlier model. In late July, a newer model found 166. The October run found 948.

The whole October run cost $1,213,430 in fees for the AI. That works out to $1,279 for each working attack. OpenAI compares this with past competition prizes of $20,000 to $70,000 for a single attack of this kind. By that measure, the work paid for itself many times over.

The AI also invented at least 10 attack methods that outside experts had not seen before.

There were mistakes too. In an early run, the AI claimed 200 successful attacks. Human review found 20 of them had worked through a different flaw than the one assigned. OpenAI added an independent check after that. The lesson for any company using AI agents: check how the result was reached, as well as the result itself.

What it means for companies in Indonesia

The people who maintain Linux are moving fast. Within 10 days of the scan, 61 of the flaws were already fixed. Another 43 had fixes waiting. Many more were being discussed in public.

That speed cuts both ways. Once a flaw is public, attackers can study it. OpenAI expects that even weaker AI models may soon turn published flaws into attacks. So how fast your team applies each fix now matters a great deal.

OpenAI's own advice is the headline for any IT team. For now, a basic login on an unpatched system should be treated as the same as full control for a well-funded attacker. Think about every shared server and contractor account. Think about the old machine that nobody updates.

Indonesian banks and telcos run many Linux servers, and so do online stores. Many leave older servers untouched because they still work. Those are exactly the machines this research puts at risk.

What to do next

• Ask your IT team how long it takes to apply a Linux security fix today, from release to every server.
• Find the machines nobody patches, such as old servers and contractor systems, and set a date for each.
• Remove basic logins that no longer need to exist, since each one is now a possible way in.
• Keep important systems separated, so one compromised machine can't reach the rest.
• If you use AI agents for security work, add a person or a separate check that confirms how each result was found.