[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-post-\u002Fnews\u002Fopenai-cyber-model-948-linux-privilege-escalation":3,"news-related-en-\u002Fnews\u002Fopenai-cyber-model-948-linux-privilege-escalation":186,"hreflang-openai-cyber-948-linux-lpe-2026-11":206},{"doc":4,"redirect":8},{"id":5,"title":6,"author":7,"author_avatar":8,"body":9,"category":166,"date":167,"description":168,"extension":169,"image":170,"meta":171,"navigation":172,"path":173,"seo":174,"seoDescription":175,"seoTitle":176,"stem":177,"tags":178,"translationKey":184,"updatedAt":8,"__hash__":185},"news_en\u002Fnews\u002Fopenai-cyber-model-948-linux-privilege-escalation.md","OpenAI's upcoming cyber model built 948 Linux root exploits for $1.2M","Paul Lefizelier",null,{"type":10,"value":11,"toc":156},"minimark",[12,25,30,38,73,76,80,88,94,98,105,112,116,122,128,132,138,142,148],[13,14,15,16,24],"p",{},"On 4 November 2026, OpenAI published ",[17,18,23],"a",{"href":19,"rel":20,"target":22},"https:\u002F\u002Fopenai.com\u002Findex\u002Fa-thousand-fewer-paths-to-root\u002F",[21],"nofollow","_blank","\"A thousand fewer paths to root\"",", a write-up of an autonomous pipeline that looks for Linux kernel vulnerabilities, proves they can be exploited and drafts patches. Powered by what OpenAI calls \"an upcoming cyber model\", one run produced 948 unique local privilege escalation (LPE) vulnerabilities, bugs that let an unprivileged local user become root, at an API cost of $1.2 million.",[26,27,29],"h2",{"id":28},"the-mid-october-run-in-numbers","The mid-October run in numbers",[13,31,32,33,37],{},"All figures below come from ",[17,34,36],{"href":19,"rel":35,"target":22},[21],"OpenAI's post",":",[39,40,41,49,55,61,67],"ul",{},[42,43,44,48],"li",{},[45,46,47],"strong",{},"Targets:"," Scout agents picked 5,034 audit targets, about 2.8 million lines of code, roughly 7% of the kernel.",[42,50,51,54],{},[45,52,53],{},"Candidates:"," analyst agents produced 2,308 candidate findings after deduplication, within 12 hours.",[42,56,57,60],{},[45,58,59],{},"Reproduced:"," 1,724 vulnerabilities were dynamically reproduced in isolated virtual machines, about 74.7% of candidates. OpenAI says only 3.1% of findings reflected incorrect analysis by the model.",[42,62,63,66],{},[45,64,65],{},"Exploited:"," agents produced 291 information leak proofs of concept and 1,169 working LPE exploits. After variant analysis, OpenAI wrote 948 distinct patches and uses that number as the count of unique LPEs.",[42,68,69,72],{},[45,70,71],{},"Cost:"," $1,213,430 in API credits in total, or $1,279 per LPE exploit.",[13,74,75],{},"The target machines ran \"a stock desktop operating system, not a weakened target\", with standard protections such as SMEP, SMAP and KASLR enabled. OpenAI says it focused on capability rather than cost, so the reported costs are \"more an upper bound\".",[26,77,79],{"id":78},"how-the-pipeline-works","How the pipeline works",[13,81,82,83,87],{},"OpenAI describes five stages: vulnerability discovery, dynamic validation, exploitability assessment, patch generation, and human review. For each finding the system bundles a root cause analysis, a minimal crashing proof of concept, a candidate patch and a full exploit. Security researchers review the exploit, patch and proposed severity before anything reaches maintainers (",[17,84,86],{"href":19,"rel":85,"target":22},[21],"OpenAI",").",[13,89,90,91,87],{},"One early incident shows why that last step matters. In an early run, 20 of 200 \"successful\" exploits had reached root through a different zero-day than the one the agent was assigned. OpenAI then added an independent check that each exploit used its assigned vulnerability (",[17,92,86],{"href":19,"rel":93,"target":22},[21],[26,95,97],{"id":96},"bounty-maths-openais-own-hypothetical","Bounty maths: OpenAI's own hypothetical",[13,99,100,101,104],{},"OpenAI frames the cost against public bug bounties. It says Pwn2Own and TyphoonPWN historically offered $20,000 and $70,000 for Linux kernel LPE exploits. Applying those amounts to the 948 exploits gives what OpenAI calls \"a hypothetical valuation of $18.9M—$66.3M, approximately 16x to 55x our pipeline's cost\" (",[17,102,86],{"href":19,"rel":103,"target":22},[21],"). That is OpenAI's estimate, not money paid.",[13,106,107,108,111],{},"A second test targeted Google's kernelCTF, a hardened kernel with a reduced attack surface. Over 112 hours and roughly $32,936 in API credits, the pipeline produced 25 information leak proofs of concept and 21 LPE exploits. For comparison, OpenAI notes kernelCTF has accepted 26 confirmed zero-day LPE exploits in 2026 so far. If all 21 were accepted, OpenAI estimates a bounty of around $1.5M, about 45x the cost (",[17,109,86],{"href":19,"rel":110,"target":22},[21],"). Its conclusion: based on public payouts, the pipeline \"was net profitable at the time of this research\".",[26,113,115],{"id":114},"newer-models-find-much-more","Newer models find much more",[13,117,118,119,87],{},"The jump between model generations is the striking part. In June, as part of OpenAI's Patch the Planet initiative, GPT-5.5 Cyber had developed 24 LPE exploits. In late July, GPT-5.6 Sol found and exploited 166 unique LPEs for $513,690. The mid-October run with the newer cyber model reached 948, on the latest kernel master branch in both cases (",[17,120,86],{"href":19,"rel":121,"target":22},[21],[13,123,124,125,87],{},"OpenAI also says that, to the best of its knowledge, the agents used at least 10 new Linux exploitation techniques, which outside kernel exploitation experts had not seen either. It withholds the technical details \"for safety reasons\" (",[17,126,86],{"href":19,"rel":127,"target":22},[21],[26,129,131],{"id":130},"patches-and-maintainers","Patches and maintainers",[13,133,134,135,87],{},"OpenAI says it has already sent hundreds of reports and patches for critical vulnerabilities, following guidance from the Linux kernel community, and is working with Patch the Planet partners on the rest, as part of its Daybreak commitment to open-source projects. Maintainers moved fast: when the 10-day scan and analysis finished, 61 LPEs were already patched upstream, 43 had staging patches and 299 were being discussed publicly (",[17,136,86],{"href":19,"rel":137,"target":22},[21],[26,139,141],{"id":140},"why-it-matters","Why it matters",[13,143,144,145,87],{},"OpenAI's practical advice is blunt: for the near and medium term, a low-privileged account on an operating system that is not up to date on patches \"should be considered functionally equivalent to root access for a well-resourced threat actor\". It recommends sandboxing and isolation such as Firecracker and gVisor to limit the damage (",[17,146,86],{"href":19,"rel":147,"target":22},[21],[13,149,150,151,155],{},"For anyone running AI coding agents on shared machines or CI runners, the takeaway is simple: keep kernels patched and treat local access as a serious boundary. OpenAI is also paying outsiders to find flaws in its own products, as covered in our piece on its ",[17,152,154],{"href":153},"\u002Fnews\u002Fopenai-bug-bounty-172000-chatgpt-codex-api\u002F","bug bounty of up to $172,000",".",{"title":157,"searchDepth":158,"depth":158,"links":159},"",2,[160,161,162,163,164,165],{"id":28,"depth":158,"text":29},{"id":78,"depth":158,"text":79},{"id":96,"depth":158,"text":97},{"id":114,"depth":158,"text":115},{"id":130,"depth":158,"text":131},{"id":140,"depth":158,"text":141},"ai","2026-11-05T12:30:00","On 4 November 2026 OpenAI described an autonomous pipeline that found, exploited and patched Linux kernel bugs. One mid-October run produced 948 unique local privilege escalation vulnerabilities at an API cost of $1,213,430.","md","\u002Fimages\u002Fnews\u002Fopenai-cyber-948-linux-privilege-escalation.jpg",{},true,"\u002Fnews\u002Fopenai-cyber-model-948-linux-privilege-escalation",{"title":6,"description":168},"OpenAI's autonomous pipeline reproduced 1,724 Linux kernel bugs and exploited 948 unique root escalations for $1,213,430 in API credits, about $1,279 each. Patches go to maintainers.","OpenAI cyber model: 948 Linux privilege escalations for $1.2M","news\u002Fopenai-cyber-model-948-linux-privilege-escalation",[179,180,181,182,183],"openai","cybersecurity","linux","ai-agents","vulnerabilities","openai-cyber-948-linux-lpe-2026-11","F8nU7JEd_blPwtDX5SADRf9Hyj9omtMmMhJZFg2w1mY",[187,193,200],{"title":188,"description":189,"path":190,"date":191,"image":192,"category":166},"Anthropic reportedly signs its first publisher licensing deal, with the AP","Adweek reports that Anthropic struck a content and data licensing deal with the Associated Press just before the US midterms, its first with a digital publisher. Neither company has confirmed it on the record, and the terms are undisclosed.","\u002Fnews\u002Fanthropic-first-publisher-licensing-deal-associated-press","2026-11-05T12:35:00","\u002Fimages\u002Fnews\u002Fanthropic-ap-publisher-licensing-deal.jpg",{"title":194,"description":195,"path":196,"date":197,"image":198,"category":199},"OpenAI buys Pierre, Meta buys Conductor: the race for the coding agent toolchain","On 4 November 2026 OpenAI announced it will acquire Pierre Computer, maker of the Git service Code Storage, and Meta bought Conductor, an app that runs several coding agents in parallel. Neither company disclosed a price.","\u002Fnews\u002Fopenai-pierre-computer-meta-conductor-coding-agent-acquisitions","2026-11-05T08:30:00","\u002Fimages\u002Fnews\u002Fopenai-pierre-meta-conductor-coding-agents.jpg","vibecoding",{"title":201,"description":202,"path":203,"date":204,"image":205,"category":166},"Anthropic's FlashProve aims to prove which AI model actually answered you","On 2 November 2026 Anthropic's Frontier Red Team released FlashProve, an open-source prototype that lets a provider prove a reply came from a committed model without revealing its weights. It is research code, with no deployment planned.","\u002Fnews\u002Fanthropic-flashprove-provable-inference-which-model-answered","2026-11-04T22:15:00","\u002Fimages\u002Fnews\u002Fanthropic-flashprove-provable-inference.jpg",{"en":207,"fr":311,"es":8},{"id":5,"title":6,"author":7,"author_avatar":8,"body":208,"category":166,"date":167,"description":168,"extension":169,"image":170,"meta":308,"navigation":172,"path":173,"seo":309,"seoDescription":175,"seoTitle":176,"stem":177,"tags":310,"translationKey":184,"updatedAt":8,"__hash__":185},{"type":10,"value":209,"toc":300},[210,215,217,222,244,246,248,253,258,260,265,270,272,277,282,284,289,291,296],[13,211,15,212,24],{},[17,213,23],{"href":19,"rel":214,"target":22},[21],[26,216,29],{"id":28},[13,218,32,219,37],{},[17,220,36],{"href":19,"rel":221,"target":22},[21],[39,223,224,228,232,236,240],{},[42,225,226,48],{},[45,227,47],{},[42,229,230,54],{},[45,231,53],{},[42,233,234,60],{},[45,235,59],{},[42,237,238,66],{},[45,239,65],{},[42,241,242,72],{},[45,243,71],{},[13,245,75],{},[26,247,79],{"id":78},[13,249,82,250,87],{},[17,251,86],{"href":19,"rel":252,"target":22},[21],[13,254,90,255,87],{},[17,256,86],{"href":19,"rel":257,"target":22},[21],[26,259,97],{"id":96},[13,261,100,262,104],{},[17,263,86],{"href":19,"rel":264,"target":22},[21],[13,266,107,267,111],{},[17,268,86],{"href":19,"rel":269,"target":22},[21],[26,271,115],{"id":114},[13,273,118,274,87],{},[17,275,86],{"href":19,"rel":276,"target":22},[21],[13,278,124,279,87],{},[17,280,86],{"href":19,"rel":281,"target":22},[21],[26,283,131],{"id":130},[13,285,134,286,87],{},[17,287,86],{"href":19,"rel":288,"target":22},[21],[26,290,141],{"id":140},[13,292,144,293,87],{},[17,294,86],{"href":19,"rel":295,"target":22},[21],[13,297,150,298,155],{},[17,299,154],{"href":153},{"title":157,"searchDepth":158,"depth":158,"links":301},[302,303,304,305,306,307],{"id":28,"depth":158,"text":29},{"id":78,"depth":158,"text":79},{"id":96,"depth":158,"text":97},{"id":114,"depth":158,"text":115},{"id":130,"depth":158,"text":131},{"id":140,"depth":158,"text":141},{},{"title":6,"description":168},[179,180,181,182,183],{"id":312,"title":313,"author":7,"author_avatar":8,"body":314,"category":166,"date":167,"description":456,"extension":169,"image":170,"meta":457,"navigation":172,"path":458,"seo":459,"seoDescription":460,"seoTitle":461,"stem":462,"tags":463,"translationKey":184,"updatedAt":8,"__hash__":467},"news_fr\u002Ffr\u002Fnews\u002Fopenai-modele-cyber-948-elevations-privileges-linux.md","Le futur modèle cyber d'OpenAI a produit 948 exploits root sous Linux pour 1,2 M$",{"type":10,"value":315,"toc":448},[316,324,328,336,368,371,375,381,387,391,398,405,409,415,421,425,431,435,441],[13,317,318,319,323],{},"Le 4 novembre 2026, OpenAI a publié ",[17,320,322],{"href":19,"rel":321,"target":22},[21],"« A thousand fewer paths to root »",", le compte rendu d'un pipeline autonome qui cherche des failles dans le noyau Linux, prouve qu'elles sont exploitables et rédige des correctifs. Propulsé par ce qu'OpenAI appelle « un futur modèle cyber », un seul passage a produit 948 vulnérabilités uniques d'élévation de privilèges locale (LPE), des bugs qui permettent à un utilisateur local sans droits de devenir root, pour un coût d'API de 1,2 million de dollars.",[26,325,327],{"id":326},"le-passage-de-mi-octobre-en-chiffres","Le passage de mi-octobre en chiffres",[13,329,330,331,335],{},"Tous les chiffres ci-dessous viennent du ",[17,332,334],{"href":19,"rel":333,"target":22},[21],"billet d'OpenAI"," :",[39,337,338,344,350,356,362],{},[42,339,340,343],{},[45,341,342],{},"Cibles :"," des agents « Scout » ont retenu 5 034 cibles d'audit, soit environ 2,8 millions de lignes de code, à peu près 7 % du noyau.",[42,345,346,349],{},[45,347,348],{},"Candidats :"," des agents d'analyse ont produit 2 308 découvertes candidates après dédoublonnage, en 12 heures.",[42,351,352,355],{},[45,353,354],{},"Reproduites :"," 1 724 vulnérabilités ont été reproduites dynamiquement dans des machines virtuelles isolées, environ 74,7 % des candidats. Selon OpenAI, seules 3,1 % des découvertes reposaient sur une analyse erronée du modèle.",[42,357,358,361],{},[45,359,360],{},"Exploitées :"," les agents ont produit 291 preuves de concept de fuite d'information et 1 169 exploits LPE fonctionnels. Après analyse des variantes, OpenAI a écrit 948 correctifs distincts et retient ce chiffre comme nombre de LPE uniques.",[42,363,364,367],{},[45,365,366],{},"Coût :"," 1 213 430 $ de crédits d'API au total, soit 1 279 $ par exploit LPE.",[13,369,370],{},"Les machines ciblées tournaient sur « un système de bureau standard, pas une cible affaiblie », avec les protections habituelles comme SMEP, SMAP et KASLR activées. OpenAI précise avoir mesuré les capacités plutôt qu'optimisé les coûts : les montants annoncés sont donc plutôt « une borne haute ».",[26,372,374],{"id":373},"comment-fonctionne-le-pipeline","Comment fonctionne le pipeline",[13,376,377,378,87],{},"OpenAI décrit cinq étapes : découverte des vulnérabilités, validation dynamique, évaluation de l'exploitabilité, génération de correctifs et revue humaine. Pour chaque faille, le système regroupe une analyse de cause, une preuve de concept minimale qui fait planter le noyau, un correctif candidat et un exploit complet. Des chercheurs en sécurité relisent l'exploit, le correctif et la gravité proposée avant tout envoi aux mainteneurs (",[17,379,86],{"href":19,"rel":380,"target":22},[21],[13,382,383,384,87],{},"Un incident montre pourquoi cette dernière étape compte. Lors d'un premier passage, 20 des 200 exploits « réussis » avaient obtenu root grâce à une autre faille zero-day que celle confiée à l'agent. OpenAI a ensuite ajouté une vérification indépendante pour s'assurer que chaque exploit utilise bien la vulnérabilité assignée (",[17,385,86],{"href":19,"rel":386,"target":22},[21],[26,388,390],{"id":389},"le-calcul-des-primes-une-hypothèse-dopenai","Le calcul des primes : une hypothèse d'OpenAI",[13,392,393,394,397],{},"OpenAI compare ce coût aux primes de bug bounty publiques. L'entreprise indique que Pwn2Own et TyphoonPWN ont historiquement offert 20 000 $ et 70 000 $ pour un exploit LPE du noyau Linux. Appliqués aux 948 exploits, ces montants donnent ce qu'OpenAI appelle « une valorisation hypothétique de 18,9 à 66,3 M$, environ 16 à 55 fois le coût du pipeline » (",[17,395,86],{"href":19,"rel":396,"target":22},[21],"). C'est une estimation d'OpenAI, pas de l'argent versé.",[13,399,400,401,404],{},"Un second test visait kernelCTF de Google, un noyau durci avec une surface d'attaque réduite. En 112 heures et environ 32 936 $ de crédits d'API, le pipeline a produit 25 preuves de concept de fuite d'information et 21 exploits LPE. À titre de comparaison, OpenAI note que kernelCTF a accepté 26 exploits LPE zero-day confirmés en 2026 à ce jour. Si les 21 étaient acceptés, OpenAI estime la prime à environ 1,5 M$, soit près de 45 fois le coût (",[17,402,86],{"href":19,"rel":403,"target":22},[21],"). Sa conclusion : au vu des primes publiques, le pipeline « était rentable au moment de cette recherche ».",[26,406,408],{"id":407},"les-nouveaux-modèles-trouvent-beaucoup-plus","Les nouveaux modèles trouvent beaucoup plus",[13,410,411,412,87],{},"L'écart entre générations de modèles est le point marquant. En juin, dans le cadre de l'initiative Patch the Planet d'OpenAI, GPT-5.5 Cyber avait développé 24 exploits LPE. Fin juillet, GPT-5.6 Sol a trouvé et exploité 166 LPE uniques pour 513 690 $. Le passage de mi-octobre avec le nouveau modèle cyber est monté à 948, à chaque fois sur la branche master la plus récente du noyau (",[17,413,86],{"href":19,"rel":414,"target":22},[21],[13,416,417,418,87],{},"OpenAI affirme aussi qu'à sa connaissance les agents ont utilisé au moins 10 nouvelles techniques d'exploitation Linux, inconnues également des experts externes consultés. Les détails techniques ne sont pas publiés « pour des raisons de sécurité » (",[17,419,86],{"href":19,"rel":420,"target":22},[21],[26,422,424],{"id":423},"correctifs-et-mainteneurs","Correctifs et mainteneurs",[13,426,427,428,87],{},"OpenAI dit avoir déjà envoyé des centaines de rapports et de correctifs pour les failles critiques, en suivant les recommandations de la communauté du noyau Linux, et travailler avec les partenaires de Patch the Planet pour le reste, dans le cadre de son engagement Daybreak envers les projets open source. Les mainteneurs ont été rapides : à la fin des 10 jours de scan et d'analyse, 61 LPE étaient déjà corrigées en amont, 43 avaient des correctifs en préparation et 299 étaient discutées publiquement (",[17,429,86],{"href":19,"rel":430,"target":22},[21],[26,432,434],{"id":433},"pourquoi-cest-important","Pourquoi c'est important",[13,436,437,438,87],{},"Le conseil pratique d'OpenAI est direct : à court et moyen terme, un compte peu privilégié sur un système qui n'a pas ses derniers correctifs « devrait être considéré comme fonctionnellement équivalent à un accès root pour un attaquant bien équipé ». L'entreprise recommande l'isolation et le sandboxing, avec des outils comme Firecracker et gVisor, pour limiter les dégâts (",[17,439,86],{"href":19,"rel":440,"target":22},[21],[13,442,443,444,155],{},"Pour qui fait tourner des agents de code IA sur des machines partagées ou des runners de CI, la leçon est simple : garder les noyaux à jour et traiter l'accès local comme une vraie frontière de sécurité. OpenAI paie aussi des chercheurs externes pour trouver des failles dans ses propres produits, comme nous l'avons raconté dans notre article sur son ",[17,445,447],{"href":446},"\u002Ffr\u002Fnews\u002Fopenai-bug-bounty-172000-dollars-chatgpt-codex\u002F","bug bounty jusqu'à 172 000 dollars",{"title":157,"searchDepth":158,"depth":158,"links":449},[450,451,452,453,454,455],{"id":326,"depth":158,"text":327},{"id":373,"depth":158,"text":374},{"id":389,"depth":158,"text":390},{"id":407,"depth":158,"text":408},{"id":423,"depth":158,"text":424},{"id":433,"depth":158,"text":434},"Le 4 novembre 2026, OpenAI a décrit un pipeline autonome qui trouve, exploite et corrige des failles du noyau Linux. Un passage mi-octobre a produit 948 élévations de privilèges locales uniques pour 1 213 430 $ de coût d'API.",{},"\u002Ffr\u002Fnews\u002Fopenai-modele-cyber-948-elevations-privileges-linux",{"title":313,"description":456},"Le pipeline autonome d'OpenAI a reproduit 1 724 failles du noyau Linux et exploité 948 élévations vers root pour 1 213 430 $ d'API, environ 1 279 $ chacune. Les correctifs vont aux mainteneurs.","Modèle cyber d'OpenAI : 948 élévations de privilèges Linux","fr\u002Fnews\u002Fopenai-modele-cyber-948-elevations-privileges-linux",[179,464,181,465,466],"cybersecurite","agents-ia","vulnerabilites","-qTLIS4GCTXsx8NJfBjgs8BH_5IoQEoyDct_L5660vQ"]