---
title: AI models discover over 400 Linux zero-day vulnerabilities in 2.5 hours each, with latest GPT-6 Cyber finding 948 exploitable kernel flaws
url: https://www.elseif.net/ai-models-discover-over-400-linux-zero-day-vulnerabilities-in-25-hours-each-with-latest-gpt-6-cyber-finding-948-exploitable-kernel-flaws
published: 2026-11-11T22:05:06+00:00
language: en
section: Agents
source: https://www.infoq.cn/article/6kuDZAK0IFkKviGrGzIR?utm_source=rss&utm_medium=article
organizations: OpenAI, Cloud Security Alliance, Databricks, Robinhood
publisher: elseif
---

# AI models discover over 400 Linux zero-day vulnerabilities in 2.5 hours each, with latest GPT-6 Cyber finding 948 exploitable kernel flaws

Researchers at OpenAI conducted an experiment in which they tasked an AI model with privilege escalation on Linux. The model discovered more than 400 previously unknown vulnerabilities allowing escalation from local user to root access, producing executable binaries that granted root privileges immediately upon execution. Each exploit required an average of two and a half hours from initial research through identification, testing, verification, and delivery of the working code. Senior reverse engineers at the presentation required weeks to understand how the AI had located some of these vulnerabilities.

In a more recent development, OpenAI's GPT-6 Cyber model developed working privilege escalation exploits for 948 Linux kernel vulnerabilities. The model reproduced 1724 vulnerabilities and successfully exploited 948 of them, representing more than five times the success rate of the previous generation model. The API costs for this work totaled approximately 1.21 million dollars, averaging 1279 dollars per successful exploit. At least 10 of the exploitation techniques employed had never been publicly disclosed before.

Caleb Sima, formerly security lead at Databricks and chief security officer at Robinhood, now chairs the Cloud Security Alliance AI Security Initiative. Two years ago, Sima discussed AI capabilities with leading red team members and reverse engineers on his podcast, and they reached consensus that AI could not perform genuine penetration testing. Their assessment held that the true value of red team work lay in connecting seemingly unrelated elements, something they believed AI could not accomplish. Sima acknowledged that this assessment proved entirely incorrect.

Historically, attackers faced economic constraints when targeting ordinary enterprises. Researching open source components used by a company and discovering a zero-day vulnerability required costs that were economically unjustifiable except in cases involving substantial bug bounties. The threat landscape has shifted fundamentally. AI systems now simultaneously scan websites, research existing vulnerabilities, discover zero-day exploits, and conduct phishing campaigns against employee targets. Once inside a network, these systems can extract credentials from developer tool configurations and use stolen keys and sessions to operate within victim systems. By the time security personnel begin reverse engineering the malicious code left behind, the AI has already begun lateral movement within the internal network.

Recently, an OpenSSH vulnerability hidden for a decade was discovered through the use of an AI Agent that conducted approximately 17600 authentication tests to expose the previously overlooked vulnerability path. The flaw existed at the intersection of OpenSSH, musl, and account configuration, affecting specific environments including Alpine Linux. Linux kernel components and surrounding infrastructure now face simultaneous re-examination by AI systems. Software that has operated for years and undergone repeated security reviews continues to expose vulnerabilities sufficient to allow attackers to assume complete system control.

OpenSSH maintainers warned in recent release notes that vulnerabilities discovered by AI have been independently found by other researchers multiple times, and attackers likely possess equivalent discovery capabilities. In response, they decided to accelerate release frequency to deliver fixes to users more rapidly. Historically, sufficiently obscure vulnerabilities might remain undiscovered for years. Currently, problems newly identified by defenders may simultaneously be under verification by attackers. The Linux ecosystem must address continuous exposure of security gaps as more powerful models repeatedly scan accumulated historical code.

The fundamental contradiction now facing enterprises involves speed disparity. Vulnerability discovery and exploitation now operate at machine speed, while enterprise asset inventory, patch verification, and deployment scheduling proceed at meeting speed. As long-accumulated problems in Linux environments begin systematic exposure, the time available for gradual remediation continues to shrink.