OpenAI's Autonomous AI Pipeline Uncovers 948 Unique Linux Kernel Privilege-Escalation Exploits
OpenAI disclosed research, "A thousand fewer paths to root," describing an autonomous AI pipeline that discovered, validated, exploited and patched Linux kernel local privilege escalation (LPE) flaws. In a single mid-October 2026 run costing $1,213,430 (~$1,279 per exploit), the pipeline analyzed roughly 7% of the kernel codebase and produced 948 unique, patched LPE vulnerabilities for a widely used Linux distribution, alongside 291 information-leak proofs of concept. This expands on earlier work under OpenAI's Patch the Planet initiative, where GPT-5.5 Cyber had found only 24 LPE exploits, and significantly outperforms a July GPT-5.6 run that found 166 LPEs despite three months of intervening kernel hardening. The findings underscore that AI-assisted vulnerability research can now operate at a scale and economic efficiency that outpaces current patch cycles, with OpenAI warning that any unpatched low-privileged account should be treated as equivalent to root access against a well-resourced attacker.
Detalhes técnicos
The five-stage pipeline (discovery, dynamic validation, exploitability assessment, patch generation, human review) used Scout agents to select 5,034 audit targets (~2.8 million lines of code), Analyst agents to generate 2,308 candidate findings in 12 hours, and a Triage agent running KASAN-instrumented VMs to reproduce 1,724 vulnerabilities (74.7% success) within 30 hours. Exploit-development agents bypassed SMEP, SMAP, KASLR and HARDENED_USERCOPY protections, producing 1,169 working LPE exploits later deduplicated to 948 unique bugs via variant analysis; exploitation alone cost $775,929, averaging ~$360 per success. The agents reportedly used at least 10 exploitation techniques new to third-party kernel experts, including persistent elastic objects, page-grooming methods, data-only privilege escalation, and an arbitrary type confusion — details withheld for safety. A separate 112-hour run on hardened kernelCTF cost ~$32,936 and yielded 21 LPE exploits. OpenAI also noted an earlier accuracy issue where 20 of 200 reported exploits actually triggered a different zero-day than assigned, prompting added independent verification checks.
Remediação
OpenAI recommends rapid patching of kernel LPEs, continuous rescanning of code as AI models improve, and sandboxing with tools such as Firecracker and gVisor. Upstream maintainers have already addressed 61 of the identified LPEs, with 43 having staging patches and 299 under public discussion; OpenAI has submitted hundreds of reports and patches under its Daybreak commitment and is coordinating with partners on remaining issues.
Linha do tempo
-
OpenAI's Autonomous AI Pipeline Uncovers 948 Unique Linux Kernel Privilege-Escalation Exploits
ler a análise desta fonte ocultar
OpenAI disclosed research, "A thousand fewer paths to root," describing an autonomous AI pipeline that discovered, validated, exploited and patched Linux kernel local privilege escalation (LPE) flaws. In a single mid-October 2026 run costing $1,213,430 (~$1,279 per exploit), the pipeline analyzed roughly 7% of the kernel codebase and produced 948 unique, patched LPE vulnerabilities for a widely used Linux distribution, alongside 291 information-leak proofs of concept. This expands on earlier work under OpenAI's Patch the Planet initiative, where GPT-5.5 Cyber had found only 24 LPE exploits, and significantly outperforms a July GPT-5.6 run that found 166 LPEs despite three months of intervening kernel hardening. The findings underscore that AI-assisted vulnerability research can now operate at a scale and economic efficiency that outpaces current patch cycles, with OpenAI warning that any unpatched low-privileged account should be treated as equivalent to root access against a well-resourced attacker.
Technical Details
The five-stage pipeline (discovery, dynamic validation, exploitability assessment, patch generation, human review) used Scout agents to select 5,034 audit targets (~2.8 million lines of code), Analyst agents to generate 2,308 candidate findings in 12 hours, and a Triage agent running KASAN-instrumented VMs to reproduce 1,724 vulnerabilities (74.7% success) within 30 hours. Exploit-development agents bypassed SMEP, SMAP, KASLR and HARDENED_USERCOPY protections, producing 1,169 working LPE exploits later deduplicated to 948 unique bugs via variant analysis; exploitation alone cost $775,929, averaging ~$360 per success. The agents reportedly used at least 10 exploitation techniques new to third-party kernel experts, including persistent elastic objects, page-grooming methods, data-only privilege escalation, and an arbitrary type confusion — details withheld for safety. A separate 112-hour run on hardened kernelCTF cost ~$32,936 and yielded 21 LPE exploits. OpenAI also noted an earlier accuracy issue where 20 of 200 reported exploits actually triggered a different zero-day than assigned, prompting added independent verification checks.
Remediation
OpenAI recommends rapid patching of kernel LPEs, continuous rescanning of code as AI models improve, and sandboxing with tools such as Firecracker and gVisor. Upstream maintainers have already addressed 61 of the identified LPEs, with 43 having staging patches and 299 under public discussion; OpenAI has submitted hundreds of reports and patches under its Daybreak commitment and is coordinating with partners on remaining issues.