OpenAI AI Pipeline Discovers 948 Exploitable Linux Kernel Vulnerabilities, Builds Working Root Exploits
OpenAI disclosed that an autonomous AI research pipeline, powered by an unreleased cyber model, discovered and built working root exploits for 948 unique Linux kernel vulnerabilities during a single mid-October run costing $1.2 million in API credits. The agents scanned the kernel's master branch, flagged 2,308 potential bugs, confirmed 1,724, and produced 1,169 working local privilege escalation (LPE) exploits plus 291 information-leak PoCs. OpenAI says the agents used at least 10 novel exploitation techniques unknown to outside kernel experts, details of which are being withheld for safety reasons. All findings were reviewed by OpenAI security researchers before being reported to kernel maintainers, and hundreds of patches have already been submitted through the company's Patch the Planet initiative. OpenAI separately raised bug bounty payouts (up to $172,000 for critical ChatGPT/Codex/API Platform bugs) to encourage external researchers to find issues its own testing may miss.
Entenda a falha
A vulnerabilidade
The story covers 948 distinct Linux kernel vulnerabilities, with no CVEs listed, found in the kernel's master branch (the most recent code). The articles give no per-bug details. They say the flaws are in areas reachable by low-privileged users, and that the agents produced 1,169 working local privilege escalation (LPE) exploits and 291 information-leak proofs of concept. An attacker needs a low-privileged local account and gets root. The exploits worked on unmodified desktop Linux VMs with KASLR, SMEP and SMAP enabled. OpenAI says the 948 figure is the number of separate patches created for the bugs behind those exploits. No affected or fixed versions are named; fixes are being sent to kernel maintainers in stages.
Como funciona
The articles do not describe the mechanics of individual bugs. They describe the discovery pipeline: AI agents first narrowed the kernel to security-relevant areas that low-privileged users can reach, then flagged 2,308 potential bugs and confirmed 1,724. For the confirmed bugs, the agents wrote exploits that turn a local, unprivileged foothold into root, or leak kernel information, on a system with common mitigations enabled. OpenAI says the agents used at least 10 novel exploitation techniques, but those details are withheld for safety reasons.
Por que é vulnerável
The articles do not give a root cause for the individual bugs, so none can be stated. They only say the flaws were in recent kernel master-branch code and were found despite three months of kernel hardening between the July and October scans. The fix for each bug is a separate patch, which OpenAI submits to kernel maintainers after its own researchers review the exploits and fixes.
Como foi descoberta
OpenAI says an autonomous pipeline powered by an unreleased cyber model found the bugs in a single mid-October scan that cost about $1.2 million in API credits. An earlier late-July scan with GPT-5.6 Sol yielded 166 unique LPEs. OpenAI security researchers reviewed the exploits, fixes and proposed severities before reporting to kernel maintainers. On the community's advice, OpenAI has submitted hundreds of reports and patches for critical flaws and will send the rest in stages with partners in its Patch the Planet initiative. It says that within 10 days of the scan, over 100 related LPEs were patched or had staging patches from other contributors, with about 300 more publicly discussed.
PoC existente, privada
O exploit existe, mas quem o tem não o publicou.
OpenAI built 1,169 working LPE exploits and 291 information-leak proofs of concept, validated against unmodified desktop Linux VMs with KASLR, SMEP and SMAP enabled. Against a hardened variant of Google's kernelCTF it produced 21 LPE exploits in 112 hours. The articles do not say they were published, and OpenAI is withholding the technical details of its novel techniques, so the exploits are treated as private. They are mature, working exploits, and OpenAI warns that an unpatched low-privileged account should now be considered equivalent to root against a well-resourced adversary, so kernel patching is urgent.
Explicação gerada por IA a partir da matéria e dos artigos, em 05/11/2026 16:15.
Detalhes técnicos
Exploits were validated against unmodified desktop Linux VMs with KASLR, SMEP and SMAP enabled. A July run using GPT-5.6 Sol produced 166 unique LPEs for $513,690; the October run found over five times more LPEs despite three months of intervening kernel hardening. Against a hardened variant of Google's kernelCTF, the pipeline delivered 21 LPE exploits over 112 hours for roughly $33,000. OpenAI estimates the 948 exploits would be worth $18.9–66.3 million at historical Pwn2Own/TyphoonPWN payout rates, 16–55x the scan's cost. Within 10 days of the scan, over 100 related Linux LPEs were patched or had staging patches from other contributors, with ~300 more publicly discussed.
Remediação
OpenAI warns that an unpatched low-privileged account should now be treated as functionally equivalent to root access for a well-resourced adversary, and recommends layered defenses such as sandboxing and isolation technologies (Firecracker, gVisor) to contain exploitation. Organizations should prioritize prompt Linux kernel patching as fixes are rolled out in stages by OpenAI and kernel maintainers.
Linha do tempo
-
OpenAI AI Pipeline Discovers 948 Exploitable Linux Kernel Vulnerabilities, Builds Working Root Exploits
ler a análise desta fonte ocultar
OpenAI disclosed that an autonomous AI research pipeline, powered by an unreleased cyber model, discovered and built working root exploits for 948 unique Linux kernel vulnerabilities during a single mid-October run costing $1.2 million in API credits. The agents scanned the kernel's master branch, flagged 2,308 potential bugs, confirmed 1,724, and produced 1,169 working local privilege escalation (LPE) exploits plus 291 information-leak PoCs. OpenAI says the agents used at least 10 novel exploitation techniques unknown to outside kernel experts, details of which are being withheld for safety reasons. All findings were reviewed by OpenAI security researchers before being reported to kernel maintainers, and hundreds of patches have already been submitted through the company's Patch the Planet initiative. OpenAI separately raised bug bounty payouts (up to $172,000 for critical ChatGPT/Codex/API Platform bugs) to encourage external researchers to find issues its own testing may miss.
Technical Details
Exploits were validated against unmodified desktop Linux VMs with KASLR, SMEP and SMAP enabled. A July run using GPT-5.6 Sol produced 166 unique LPEs for $513,690; the October run found over five times more LPEs despite three months of intervening kernel hardening. Against a hardened variant of Google's kernelCTF, the pipeline delivered 21 LPE exploits over 112 hours for roughly $33,000. OpenAI estimates the 948 exploits would be worth $18.9–66.3 million at historical Pwn2Own/TyphoonPWN payout rates, 16–55x the scan's cost. Within 10 days of the scan, over 100 related Linux LPEs were patched or had staging patches from other contributors, with ~300 more publicly discussed.
Remediation
OpenAI warns that an unpatched low-privileged account should now be treated as functionally equivalent to root access for a well-resourced adversary, and recommends layered defenses such as sandboxing and isolation technologies (Firecracker, gVisor) to contain exploitation. Organizations should prioritize prompt Linux kernel patching as fixes are rolled out in stages by OpenAI and kernel maintainers.