OpenAI Reports 948 Linux Security Flaws Found With AI

OpenAI says an AI system found 948 distinct ways for an ordinary user to gain administrator control in Linux tests, exposing weaknesses in software that underpins cloud infrastructure and Android devices. The findings, published November 4, show how AI could help defenders uncover serious flaws while giving attackers more opportunities if repairs fall behind. [1]

The tests targeted the Linux kernel, the core software that manages a computer’s resources. They began with access to an ordinary account and demonstrated a jump to administrator authority, known as root. Someone exploiting that kind of flaw already has a foothold on the machine; gaining greater control can put other users’ files and the system itself at risk.

OpenAI had submitted hundreds of reports and proposed fixes when it published its results. The repair work was continuing. [1] A proposed change still needs review, testing and delivery through software updates before it can protect the people running the affected code.

Security firm Trail of Bits is using AI to adapt existing repairs for older Linux releases. [2] That work addresses a separate problem: a vulnerability may be fixed in the newest code while remaining in the software an organization relies on.

Checking what the AI actually found

OpenAI’s mid-October campaign produced 1,169 working demonstrations, consolidated into 948 distinct findings by counting separate patches after examining related flaws. Charges for model use totalled about $1.21 million. These are the company’s reported results. [1]

Human review exposed a problem in an early run: 20 of 200 apparent successes exploited a different flaw from the one assigned. OpenAI added checks to confirm which vulnerability each demonstration used. [1] That distinction matters to the engineer preparing a repair: a successful takeover proves that something failed, but the fix has to address the failure that actually occurred.

OpenAI’s Patch the Planet initiative, launched with Trail of Bits in June, assigns security engineers to reproduce findings, remove duplicates, assess severity and develop patches and tests with maintainers. The company acknowledges that AI models also generate substantial numbers of false alarms. Reviewing those reports before they reach maintainers is intended to spare already stretched software teams work that produces no useful repair. Maintainers retain control over which patches they accept and how findings are disclosed. [3]

Getting repairs into older software

By October 29, Trail of Bits had helped deliver fixes for 149 distinct CVEs into stable Linux kernel releases, the firm reported November 2. CVEs are the public identifiers used to track vulnerabilities. [2]

Linux supports several release lines. New features enter the mainline version, while stable and long-term releases receive selected bug fixes. Distributors also maintain their own kernels, so a repair in the newest Linux code does not establish that every supported installation has received it. [6]

Backporting means transferring a fix to an older release. A patch may depend on newer code that the older release lacks, requiring engineers to adapt the change and check that it does not break other functions.

“We should not have to make a choice between stability and security,” wrote Trail of Bits researcher Artem Dinaburg. [2]

Trail of Bits uses Codex to find missing fixes, adapt code, run tests and prepare evidence. Engineers review the work and submit it to maintainers. [2]

Trail of Bits estimates that straightforward preparation and review now takes about 30 minutes of human attention, with more involved cases taking around three hours. Those estimates exclude initial setup, and testing for unintended failures and maintainer acceptance remain necessary. [2]

Finding a flaw before users receive it

Google’s Big Sleep researchers demonstrated an earlier opportunity for intervention in 2024, when their AI agent found a flaw in SQLite, widely used software for storing and retrieving data. Developers repaired it on the day it was reported, before it appeared in an official release. Users were spared the affected version. [5]

The defect involved memory safety: the program could write data outside an intended area of working memory. The team, a collaboration between Google Project Zero and DeepMind, had asked the agent to investigate code changes for related problems that might have escaped an earlier repair. [5]

The researchers cautioned that “these are highly experimental results.” A testing tool tailored to SQLite might have been equally effective, they said. Such tools, called fuzzers, try many inputs to uncover failures. The case showed the benefit of catching a defect during development, while leaving open which combination of tools would do that work most effectively. [5]

Attackers are adopting the tools too

Anthropic reported in September that suspected state-linked operators and financially motivated criminals had used Claude in operations it disrupted between December 2025 and August 2026. AI helped research targets, develop tools and process stolen information. Humans continued to choose targets and review stolen data. [4]

In one suspected espionage operation, Anthropic observed AI-assisted workflows that modified and rebuilt malicious tools after security products detected them. The account describes activity visible to one provider, not the prevalence of AI across cybercrime, but it shows attackers using automation throughout an operation rather than only to find new flaws. [4]

For Linux operators, the immediate response remains identifying affected systems, following their distributor’s security advisories, and testing and installing available updates. Limited account permissions and separation between sensitive systems can help contain an intrusion while repairs are deployed.

The 948 findings measure what OpenAI’s system demonstrated in its tests. The security benefit will depend on how many of those weaknesses are reliably repaired in the versions people actually run.

Primary sources

[1] OpenAI. A thousand fewer paths to root. November 4, 2026. https://openai.com/index/a-thousand-fewer-paths-to-root/

[2] Artem Dinaburg, Trail of Bits. Backporting security fixes at scale with Codex. November 2, 2026. https://blog.trailofbits.com/2026/11/02/backporting-security-fixes-at-scale-with-codex/

[3] OpenAI. Patch the Planet a Daybreak initiative to support open source maintainers. June 22, 2026. https://openai.com/index/patch-the-planet/

[4] Anthropic. Detecting and countering misuse of AI September 2026. https://www.anthropic.com/threat-intelligence-report-september-2026

[5] Google Project Zero and Google DeepMind, Big Sleep team. From Naptime to Big Sleep Using Large Language Models To Catch Vulnerabilities In Real World Code. November 1, 2024. https://projectzero.google/2024/10/from-naptime-to-big-sleep.html

[6] Linux Kernel Organization. Active kernel releases. Accessed November 5, 2026. https://www.kernel.org/category/releases.html