OpenAI's Autonomous Cyber Pipeline Finds 1,724 Linux Kernel Bugs, Builds 948 Root Exploits
OpenAI disclosed an automated vulnerability-research pipeline, powered by an upcoming OpenAI cyber model, that discovered 1,724 real Linux kernel vulnerabilities and built 948 working local privilege-escalation (root) exploits in a mid-October 2026 run against the kernel's master branch. The company used the results to generate candidate patches for human researchers to send to kernel maintainers, framing the work as a demonstration of scaled, automated vulnerability discovery, validation and remediation rather than an attack against production systems. This follows earlier OpenAI runs under the "Patch the Planet" project: GPT-5.5 Cyber found and built 24 root exploits in June 2026, and GPT-5.6 Sol found and exploited 166 unique flaws for $513,690 in July 2026. OpenAI said the agents discovered at least 10 novel Linux exploitation techniques that outside kernel exploitation experts were also unfamiliar with, and the company is keeping the full technical details private for safety reasons (securityweek.com). There is no indication in the article that any of these bugs were exploited by threat actors in the wild; they were found and reported by OpenAI itself. Subsequent reporting (thehackernews.com) placed the disclosure in the context of a broader surge in AI-assisted vulnerability research, citing independent assessments from VulnCheck, Google Threat Intelligence Group (GTIG) and the UK National Cyber Security Centre (NCSC). A later report (gbhackers.com) confirmed the same figures and detailed the research in a paper titled "A thousand fewer paths to root," noting the full pipeline cost of $1,213,430. Separately, on November 2, 2026, OpenAI announced it is significantly raising payouts in its Security Bug Bounty program to encourage more outside researchers to hunt for flaws its own testing may have missed (securityweek.com).
Entenda a falha
A vulnerabilidade
No CVEs are listed. The story covers 1,724 Linux kernel bugs that OpenAI's pipeline reproduced in the kernel's master branch, of which 948 unique ones were turned into local root exploits; 291 more PoCs only leak kernel information. The flaws sit in security-relevant code reachable by low-privileged users (syscalls, drivers, filesystems, networking). An attacker needs a local unprivileged account and gets root. Exploits worked on a stock desktop distribution with KASLR, SMEP, SMAP and HARDENED_USERCOPY. No affected or fixed versions are named. After 10 days, 61 were patched upstream, 43 had staging patches and 299 were under public discussion.
Como funciona
The articles describe the pipeline, not the mechanics of individual bugs. Scout agents picked locally reachable audit targets (about 7% of the kernel), and analyst agents produced 2,308 candidate findings. Triage agents built crashing test programs and reproduced 1,724 of them in isolated VMs running a KASAN kernel. Screening narrowed the list, and ExploitDev agents then tried to turn the bugs into info leaks or local root on machines with standard mitigations on. Unconfirmed candidates mostly needed elevated privileges or failed on race conditions and fault-injection requirements.
Por que é vulnerável
The articles give no root cause for the individual bugs, so none can be stated. They say only that the flaws were in recent master-branch code and were found despite three months of kernel hardening since the July scan. Each bug gets its own patch, generated by the pipeline and reviewed by OpenAI researchers before going to maintainers. In an early run, 20 of 200 reported exploits reached root through a different zero-day than the one assigned, so OpenAI added independent checks that each exploit uses its intended bug.
Como foi descoberta
OpenAI says an autonomous pipeline running an unreleased cyber model found the bugs in a mid-October 2026 scan that cost $1,213,430 in API credits, under its Patch the Planet initiative. Earlier runs were GPT-5.5 Cyber in June (24 LPE exploits) and GPT-5.6 Sol in July (166). OpenAI researchers review each exploit, patch and severity before reporting to kernel maintainers. At the kernel community's suggestion, hundreds of reports and patches for the most severe issues have been sent, and the rest will follow in stages with partners. Some outlets describe the 948 bugs as already patched, but the detailed counts show most fixes were still pending.
PoC existente, privada
O exploit existe, mas quem o tem não o publicou.
OpenAI built 1,169 working LPE exploits (948 unique bugs) and 291 info-leak PoCs, validated on desktop Linux VMs with KASLR, SMEP, SMAP and HARDENED_USERCOPY. On a hardened kernelCTF variant it produced 21 LPE exploits in 112 hours. The articles do not say the exploits were released, and the technique details are withheld for safety, so they are treated as private. They are mature and working, and OpenAI warns that an unpatched system with a low-privileged account should be treated as equivalent to root against a well-resourced attacker, so kernel patching is urgent.
Explicação gerada por IA a partir da matéria e dos artigos, em 05/11/2026 19:08.
Detalhes técnicos
The pipeline used a five-stage workflow: Scout agents narrowed the 40-million-line kernel to 5,034 locally-reachable targets (2.8 million lines, ~7% of the kernel) for $3,426; Analyst agents, using a workflow similar to the Codex Security plugin (gbhackers.com), produced 2,308 unique candidate findings for $316,549 within 12 hours; Triage agents built crashing test programs under Kernel Address Sanitizer in isolated VMs and dynamically reproduced 1,724 vulnerabilities (a 74.7% confirmation rate, only 3.1% of failures due to incorrect model analysis) for $73,330 within 30 hours; a screening step narrowed these to 313 info-leak and 1,238 root candidates; and ExploitDev agents built working exploits against desktop VMs hardened with KASLR, SMEP, SMAP and HARDENED_USERCOPY, producing 291 info-leak PoCs and 1,169 root exploits that, after variant analysis, mapped to 948 unique vulnerabilities. The entire run cost $1,213,430 in API credits ($1,279 per unique root exploit); exploitation alone accounted for $775,929 of that total, with successful attempts averaging about $360 each while 69 failed attempts cost far more, mainly due to hard race conditions (gbhackers.com). For comparison, Pwn2Own and TyphoonPWN pay $20,000-$70,000 per kernel root exploit, implying a hypothetical bounty value of $18.9-$66.3 million for the haul. A smaller 112-hour run against an adapted Google kernelCTF target cost $32,936 and yielded 25 info leaks and 21 root exploits, against a kernelCTF program that had accepted 26 zero-day root exploits in 2026 at the time of writing. Human reviewers caught a validation flaw early on: 20 of 200 reported "working" exploits actually gained root via a different bug than the one assigned, prompting the team to add strict path checks. Researchers sampling 166 exploits identified at least 10 novel Linux kernel exploitation techniques, including two new elastic cache-spraying objects, two page-grooming tricks, five data-only root attacks, one type-confusion method, and a crash found and exploited in the QEMU TCG emulator by a system-call auditing agent.
Remediação
OpenAI reported that outside Linux kernel contributors patched or staged fixes for over 100 flaws during the 10-day scan window, specified as 61 LPEs fixed upstream and 43 with staging patches, with 299 more under public discussion (gbhackers.com); it noted that overall, over 100 Linux LPEs had been patched or had staging patches from other contributors within 10 days of the scan, with roughly 300 more publicly discussed (securityweek.com). Defenders are advised to apply kernel updates promptly, since the authors warned that an unpatched local user account is now functionally equivalent to root access for a well-resourced threat actor, and to use strict sandboxing technologies such as Firecracker or gVisor to contain potential local-privilege-escalation exploitation. OpenAI stated it has already submitted hundreds of reports and patches as part of its Daybreak commitment and is collaborating with partners on remaining issues, and pointed to a planned Defense Factory for continuous, agent-driven remediation (gbhackers.com). OpenAI additionally stated that AI-generated patches, produced with human expert-guided iteration, can be functionally correct, aligned with existing coding styles and robust to attack variants, arguing that automated patch generation must keep pace with AI-accelerated vulnerability discovery (thehackernews.com). This broader trend is corroborated by independent sources: VulnCheck categorized AI-assisted vulnerability discovery as a "growing risk category" that lowers the technical bar for novice hackers; GTIG data shows monthly vulnerability disclosures more than doubling from 5,045 in January 2026 to 10,740 in August 2026, with high-risk disclosures up 167% over the same period; and the UK NCSC, in an assessment published 3 November 2026, judged it almost certain that AI-powered cyber tools have significantly increased zero-day discovery, while noting that defensive AI tools have shortened the zero-day-to-n-day window and sped up patching, though many systems remain exposed where the technology is not correctly and widely configured (thehackernews.com). Separately, OpenAI raised its Security Bug Bounty payouts on November 2, 2026: the midpoint of payout ranges for medium-, high- and critical-severity bugs is now more than four times higher on average, critical ChatGPT/Codex/API Platform flaws can earn up to $30,000, exceptional critical findings up to $172,000 with bonuses, and the program now includes a five-tier researcher loyalty program and extra multipliers for priority campaigns, aimed at getting more outside researchers hunting for flaws OpenAI's own testing may have missed (securityweek.com).
Linha do tempo
-
OpenAI's AI Pipeline Builds 948 Root Exploits for Linux Kernel Bugs; Firm Quadruples Bug Bounty Payouts
This article adds that kernel exploitation experts outside OpenAI were unfamiliar with the novel exploitation techniques the pipeline discovered, reinforcing why OpenAI is withholding technical details, and reports a separate November 2, 2026 announcement in which OpenAI substantially raised its Security Bug Bounty payouts — quadrupling the midpoint of medium/high/critical reward ranges, allowing critical ChatGPT/Codex/API flaws to earn up to $30,000, exceptional findings up to $172,000 with bonuses, and introducing a five-tier loyalty program and priority-campaign multipliers — a move explicitly framed as inviting outside researchers to find flaws the company's own automated testing may have missed.
ler a análise desta fonte ocultar
This article adds that kernel exploitation experts outside OpenAI were unfamiliar with the novel exploitation techniques the pipeline discovered, reinforcing why OpenAI is withholding technical details, and reports a separate November 2, 2026 announcement in which OpenAI substantially raised its Security Bug Bounty payouts — quadrupling the midpoint of medium/high/critical reward ranges, allowing critical ChatGPT/Codex/API flaws to earn up to $30,000, exceptional findings up to $172,000 with bonuses, and introducing a five-tier loyalty program and priority-campaign multipliers — a move explicitly framed as inviting outside researchers to find flaws the company's own automated testing may have missed.
OpenAI disclosed that an autonomous pipeline powered by an unreleased cyber model built working root exploits for 948 unique Linux kernel vulnerabilities in a single mid-October 2026 run, after flagging 2,308 potential bugs and confirming 1,724 of them. The run cost $1.2 million in API credits and produced 1,169 working local privilege escalation (LPE) exploits plus 291 information-leak proofs of concept, tested against hardened desktop Linux VMs with KASLR, SMEP and SMAP enabled. OpenAI says the agents discovered at least 10 novel Linux exploitation techniques that even outside kernel exploitation experts were unfamiliar with, and it is withholding the technical details for safety reasons. Separately, on November 2, OpenAI announced it is significantly raising payouts in its Security Bug Bounty program to attract more outside researchers.
Technical Details
The mid-October scan targeted the most recent code in the kernel's master branch. AI agents first narrowed the kernel down to security-relevant areas that low-privileged users can reach, then flagged 2,308 potential bugs and confirmed 1,724 of them. In total, the agents came up with 1,169 working local privilege escalation (LPE) exploits, along with 291 proofs of concept for information leaks. OpenAI said the pipeline was also pointed at an adapted version of Google's kernelCTF, a hardened target with less attack surface, delivering 21 LPE exploits over 112 hours and about $33,000 in API credits. OpenAI puts the haul in perspective: at the $20,000 and $70,000 that Pwn2Own and TyphoonPWN have historically paid for Linux kernel privilege escalation exploits, the 948 bugs would be worth $18.9 million to $66.3 million, or 16 to 55 times what the scan cost.
Remediation
Acting on advice from the kernel community, OpenAI says it has already submitted hundreds of reports and patches for critical flaws, and will send fixes for the remaining bugs in stages, working with partners in its Patch the Planet initiative. OpenAI warned that having access to a low-privileged account on an operating system that isn't up to date on its latest available patches should be considered functionally equivalent to root access for a well-resourced threat actor. To contain the damage when a flaw does get exploited, OpenAI points to layered defenses, including sandboxing and isolation technologies such as Firecracker and gVisor. Separately, the company increased Security Bug Bounty payouts so that the midpoint of its payout ranges for medium-, high- and critical-severity bugs is now more than four times higher on average, with a researcher who reports a critical flaw in ChatGPT, Codex, the API Platform or another core product able to earn as much as $30,000, and bonuses that can now push a single payout to $172,000, plus a five-tier loyalty program and extra multipliers for priority campaigns.
-
OpenAI's Cyber AI Pipeline Builds 948 Linux Root Exploits in Mid-October 2026 Run, Costing $1.2M
This article adds operational detail to the known case: it specifies that Analyst agents used a workflow similar to the Codex Security plugin, breaks down exploitation costs ($775,929 total, ~$360 per successful exploit, with 69 failed attempts costing more due to hard race conditions), and gives a finer breakdown of remediation outcomes during the scan window (61 LPEs patched upstream, 43 with staging patches, 299 under public discussion) versus the previously reported aggregate of "over 100" fixed/staged and "roughly 300" discussed. It also names OpenAI's **Daybreak commitment** for submitting reports/patches and its planned **Defense Factory** for continuous agent-driven remediation, neither of which appeared in the known report.
ler a análise desta fonte ocultar
This article adds operational detail to the known case: it specifies that Analyst agents used a workflow similar to the Codex Security plugin, breaks down exploitation costs ($775,929 total, ~$360 per successful exploit, with 69 failed attempts costing more due to hard race conditions), and gives a finer breakdown of remediation outcomes during the scan window (61 LPEs patched upstream, 43 with staging patches, 299 under public discussion) versus the previously reported aggregate of "over 100" fixed/staged and "roughly 300" discussed. It also names OpenAI's Daybreak commitment for submitting reports/patches and its planned Defense Factory for continuous agent-driven remediation, neither of which appeared in the known report.
OpenAI described an autonomous pipeline that discovers, validates, exploits and patches flaws in the Linux kernel, expanding its earlier "Patch the Planet" work. In a single run in mid-October 2026, the upcoming cybersecurity model generated 948 working local privilege-escalation (LPE) exploits for a widely used Linux distribution, at a total cost of $1,213,430 (about $1,279 per exploit), detailed in a report titled "A thousand fewer paths to root." The pipeline runs in five stages — discovery, dynamic validation, exploitability assessment, patch generation and human review — and the article is a research/disclosure report, not a record of a real-world intrusion or victim.
Technical Details
Scout agents selected 5,034 audit targets (about 2.8 million lines, roughly 7% of the kernel), and Analyst agents, using a workflow similar to the Codex Security plugin, produced 2,308 candidate findings within 12 hours. A triage agent reproduced 1,724 vulnerabilities (74.7% of candidates) in a KASAN-enabled VM within 30 hours, with only 3.1% of failures due to incorrect model analysis. Exploit-development agents then worked against SMEP, SMAP, KASLR and HARDENED_USERCOPY protections, producing 291 information-leak PoCs and 1,169 working LPE exploits that variant analysis reduced to 948 unique vulnerabilities with patches; exploitation alone accounted for $775,929 of the total cost, with successful attempts averaging about $360 each while 69 failed attempts cost far more, mainly due to hard race conditions. A separate 112-hour run on a hardened kernelCTF environment cost about $32,936 and yielded 21 LPE exploits, against 26 accepted 0-day LPEs in the kernelCTF program in 2026 so far. The agents reportedly used at least 10 exploitation techniques — including new persistent elastic objects, page-grooming methods, data-only privilege-escalation techniques and an arbitrary type confusion — that third-party kernel experts had not previously seen; OpenAI withheld technical details for safety reasons. Validation controls were tightened after an earlier run in which 20 of 200 reported exploits had achieved root through a different zero-day than the one assigned.
Remediation
During the scan window, upstream maintainers addressed 61 LPEs, another 43 had staging patches, and 299 more were under public discussion; OpenAI said it has already submitted hundreds of reports and patches as part of its Daybreak commitment and is collaborating with partners on the remaining issues. OpenAI's guidance is that an unpatched system with a low-privileged account should be treated as equivalent to root access for a well-resourced attacker, and it recommends rapid patching, rescanning code as models improve, and sandboxing with tools like Firecracker and gVisor; the company's planned Defense Factory points toward continuous, agent-driven remediation.
-
OpenAI's AI Pipeline Generates 948 Linux Root Privilege Escalation Exploits in 30-Hour Run
This article republishes OpenAI's own account of the pipeline's architecture, costs and novel exploitation techniques, and adds broader context missing from earlier reporting: VulnCheck's framing of AI-assisted vulnerability discovery as a growing risk category for novice attackers, GTIG data showing monthly vulnerability disclosures more than doubling and high-risk disclosures growing 167% between January and August 2026, and a new NCSC assessment (published 3 November 2026) stating that AI-powered cyber tools have almost certainly driven a significant rise in zero-day discovery while also shortening the zero-day-to-n-day patching window, though many systems remain exposed where the technology is not properly configured.
ler a análise desta fonte ocultar
This article republishes OpenAI's own account of the pipeline's architecture, costs and novel exploitation techniques, and adds broader context missing from earlier reporting: VulnCheck's framing of AI-assisted vulnerability discovery as a growing risk category for novice attackers, GTIG data showing monthly vulnerability disclosures more than doubling and high-risk disclosures growing 167% between January and August 2026, and a new NCSC assessment (published 3 November 2026) stating that AI-powered cyber tools have almost certainly driven a significant rise in zero-day discovery while also shortening the zero-day-to-n-day patching window, though many systems remain exposed where the technology is not properly configured.
OpenAI detailed the five-stage agentic pipeline it used in October 2026 to discover, validate and patch Linux kernel vulnerabilities, confirming the 948 working local privilege escalation (LPE) exploits and 1,724 reproduced vulnerabilities already reported, and placing the effort in a wider industry context of AI-accelerated vulnerability discovery. The article adds that the pipeline analyzed 2,308 candidate kernel findings and cost $1,213,430 in API credits, or $1,279 per exploit, building on 24 LPE exploits OpenAI's GPT-5.5-Cyber model found in June 2026. It situates the disclosure alongside independent assessments from VulnCheck, Google Threat Intelligence Group (GTIG) and the UK National Cyber Security Centre (NCSC) on how AI tools are reshaping the balance between offensive and defensive vulnerability research.
Technical Details
OpenAI said: "Our prototype identified and analyzed 2,308 candidate Linux kernel findings, dynamically reproduced 1,724 vulnerabilities, and produced 948 local privilege escalation (LPE) exploits that could allow an unprivileged local user to gain root access." The pipeline comprises five stages: vulnerability discovery, where agents identify high-value audit targets reachable by low-privileged users; dynamic validation, where agents reproduce candidates in a sandboxed environment; exploitability assessment, testing whether vulnerabilities enable information disclosure or LPE and building a proof-of-concept; patch generation; and human review by security researchers before reporting to maintainers. In the mid-October 2026 run, the pipeline dynamically reproduced 1,724 vulnerabilities within 30 hours and generated PoCs for 291 information leak vulnerabilities and 948 working LPE exploits for a commonly used Linux distribution, at a total cost of $1,213,430 in API credits, or $1,279 per LPE exploit. The agents in this pipeline discovered and used at least 10 novel Linux exploitation techniques, including two new persistent elastic objects to spray objects inside the target kernel cache, two new page grooming techniques to improve exploit reliability, five new data-only techniques to obtain privilege escalation directly or via an arbitrary write, and one technique involving arbitrary type confusion.
Remediation
OpenAI argued that the rate of AI-driven vulnerability discovery must be matched by automated patch generation, stating that "with human expert-guided iteration, we believe AI-generated patches can be functionally correct, aligned with existing coding styles, and robust to attack variants." It warned that "for the near and medium term, having access to a low-privileged account on an operating system that isn't up-to-date on its latest available patches should be considered functionally equivalent to root access for a well-resourced threat actor." Separately, VulnCheck categorized AI-assisted vulnerability discovery as a "growing risk category" that turbocharges novice hackers, who now require little to no technical expertise to detect software vulnerabilities. According to GTIG, the number of vulnerabilities disclosed per month has more than doubled, rising from 5,045 in January 2026 to 10,740 in August 2026, and vulnerabilities classified as high-risk surged from 131 disclosures in January 2026 to 350 in August 2026, a 167% growth. The NCSC assessed that "the widespread availability of AI-powered cyber tools has almost certainly led to a significant increase in the discovery of zero-day vulnerabilities." It added that defensive AI-cyber tools have reduced the window of a vulnerability moving from zero-day to an N-day vulnerability and made patching quicker, though without correct and widespread configuration many systems remain vulnerable to AI-cyber tools identifying and exploiting weaknesses.
-
OpenAI's Autonomous Cyber Pipeline Finds 1,724 Linux Kernel Bugs, Builds 948 Root Exploits
ler a análise desta fonte ocultar
OpenAI disclosed an automated vulnerability-research pipeline, powered by an upcoming OpenAI cyber model, that discovered 1,724 real Linux kernel vulnerabilities and built 948 working local privilege-escalation (root) exploits in a mid-October 2026 run against the kernel's master branch. The company used the results to generate candidate patches for human researchers to send to kernel maintainers, framing the work as a demonstration of scaled, automated vulnerability discovery, validation and remediation rather than an attack against production systems. This follows earlier OpenAI runs under the "Patch the Planet" project: GPT-5.5 Cyber found and built 24 root exploits in June 2026, and GPT-5.6 Sol found and exploited 166 unique flaws for $513,690 in July 2026. There is no indication in the article that any of these bugs were exploited by threat actors in the wild; they were found and reported by OpenAI itself.
Technical Details
The pipeline used a five-stage workflow: Scout agents narrowed the 40-million-line kernel to 5,034 locally-reachable targets (2.8 million lines, ~7% of the kernel) for $3,426; Analyst agents produced 2,308 unique candidate findings for $316,549; Triage agents built crashing test programs under Kernel Address Sanitizer in isolated VMs and dynamically reproduced 1,724 vulnerabilities (a 74.7% confirmation rate) for $73,330; a screening step narrowed these to 313 info-leak and 1,238 root candidates; and ExploitDev agents built working exploits against desktop VMs hardened with KASLR, SMEP, SMAP and HARDENED_USERCOPY, producing 291 info-leak PoCs and 1,169 root exploits that, after variant analysis, mapped to 948 unique vulnerabilities. The entire run cost $1,213,430 in API credits ($1,279 per unique root exploit, $359 per exploit built from an already-confirmed bug). For comparison, Pwn2Own and TyphoonPWN pay $20,000-$70,000 per kernel root exploit, implying a hypothetical bounty value of $18.9-$66.3 million for the haul. A smaller 112-hour run against an adapted Google kernelCTF target cost $32,936 and yielded 25 info leaks and 21 root exploits, against a kernelCTF program that had accepted 26 zero-day root exploits in 2026 at the time of writing. Human reviewers caught a validation flaw early on: 20 of 200 reported "working" exploits actually gained root via a different bug than the one assigned, prompting the team to add strict path checks. Researchers sampling 166 exploits identified at least 10 novel Linux kernel exploitation techniques, including two new elastic cache-spraying objects, two page-grooming tricks, five data-only root attacks, one type-confusion method, and a crash found and exploited in the QEMU TCG emulator by a system-call auditing agent.
Remediation
OpenAI reported that outside Linux kernel contributors patched or staged fixes for over 100 flaws during the 10-day scan window, with roughly 300 more discussed in public forums. Defenders are advised to apply kernel updates promptly, since the authors warned that an unpatched local user account is now functionally equivalent to root access for a well-resourced threat actor, and to use strict sandboxing technologies such as Firecracker or gVisor to contain potential local-privilege-escalation exploitation.