On November 5, 2025, OpenAI's security research team publicly demonstrated a "demonstration-level" automated cybersecurity experiment: a five-stage process system paired with an upcoming cybersecurity model that conducted a fully automated vulnerability mining on the Linux kernel in mid-October 2025. In a single run, this process first filtered out 2,308 suspected samples during the vulnerability discovery phase, then confirmed 1,724 of them as real exploitable vulnerabilities through dynamic reproduction, and in the feasibility assessment, patch generation, and manual review stages, it further automatically generated usable exploit code and candidate repair patches for 948 local privilege escalation vulnerabilities and produced PoCs for 291 information leakage vulnerabilities. The Linux kernel is a core component of global servers, cloud infrastructure, and many Android devices, and this operation, which "mined over a thousand kernel vulnerabilities at once," was clearly packaged as a showcase of AI security capabilities and technical disclosure, rather than a traditional intrusion report or incident notification. Public reports indicate that the cost for a single API run was about 1.21 million USD, revealing the leap in AI vulnerability mining efficiency while firmly pinning such capabilities to the high threshold of costly computing power and funding, making the dual-use risk of "efficient automated exploitation tools being able to patch systems while also being used to attack systems," as well as who qualifies and under what rules gets to hold it, the core contradiction in all future discussions.
Pipeline-style AI offense and defense: How 1,724 vulnerabilities were forged
This chain looks more like an automated production line rather than a moment of inspiration from a genius hacker. The first step is "rough screening": the model conducts large-scale static and semantic analysis on the Linux kernel. According to a single source, one run generated 2,308 suspected vulnerability candidate samples. Next comes the second phase—dynamic validation, where the system constructs inputs for each candidate point, trying to trigger anomalies, turning paper-based guesses into observable behaviors with a repeatable "one pass counts as one" approach. By the end of this phase, only 1,724 vulnerabilities confirmed as "really triggerable" remained, while the remaining samples were rejected, which constitutes the most critical reduction valve in the entire production line.
However, what is most unsettling and best reflects the "integration of offense and defense" are the subsequent three steps. In the feasibility assessment phase, AI no longer is satisfied with "having vulnerabilities," but labels each vulnerability with "whether it is worth further development," prioritizing those suitable for writing exploit codes for local privilege escalation and information leakage; ultimately generating PoCs or Exploits for 948 local privilege escalation vulnerabilities and 291 information leakage vulnerabilities, thereby materializing potential attack paths. In the subsequent patch generation phase, the model proposed candidate repair plans for these confirmed vulnerabilities, essentially simulating an attacker writing exploits while simultaneously drafting patch proposals for the defensive side. The manual review at the end of the link is the only step that "pulls back to human involvement": researchers need to review and adjust the vulnerability conclusions and patches. However, current public information does not clarify how many results the review covered, nor does it disclose the overall false positive rate, which means we know humans are still in the brake position but not how much pressure is applied.
1.21 million USD per round: The computing power threshold of AI security experiments
The cost of running the entire link once has been summarized into a single figure by several reports: about 1.21 million USD for API calls. The brief also emphasizes that this only counts the calls and does not include the infrastructure, engineering support, and manual review involved in model inference. Adding the premise of "paired with an upcoming cybersecurity model," one can read an implied condition: today, to let the model run through the entire "discovery-validation-exploitation-patching-review" stack process on targets like the Linux kernel, it still requires extremely expensive, computation-intensive models as driving forces. In other words, this experiment that uncovered 1,724 kernel vulnerabilities at once is currently only practically meaningful for a handful of institutions with sufficiently ample budgets and computing power; the security community continues to express ongoing concern regarding the "1.21 million USD computing power threshold" itself.
Yet, from a different perspective, compared with traditional security research, this amount of money is no longer merely "burning cash to showcase capabilities." In the past, finding even a few dozen reproducible vulnerabilities in a complex kernel often required researchers to analyze code line by line, repeatedly write PoCs, and debug environments, with time measured in months or even years, far from achieving the scale of confirming 1,724 vulnerabilities at once, automatically generating exploit codes and PoCs for 948 local privilege escalation vulnerabilities and 291 information leakage vulnerabilities. Even though 1.21 million USD is astonishing in absolute terms, it purchases a previously non-existent efficiency curve: outsourcing parts of work previously bottlenecked by human effort to large models for "wholesale" mining. The brief simultaneously notes that such AI vulnerability research, due to high costs, remains a high-threshold experiment in the short term, not easily replicable or abused, but as model efficiency increases and calling costs drop, this capability is likely to evolve from a few "affordable" demonstration projects to attack and defense tools that more institutions can bear, thereby gradually lowering the computing power threshold and redistributing risks more widely.
948 privilege escalation exploit codes: Invisible bombs in the Linux world
In the vocabulary of security engineers, "local privilege escalation" sounds very academic, but in reality, it is extremely direct: it allows a process that is originally just a normal account, even one with almost no permissions, to progressively climb to root-level highest permissions. Public reports indicate that local privilege escalation vulnerabilities can allow unauthorized or low-privileged users to gain complete control over the system, with users, processes, and data that were originally isolated by permission boundaries being connected through an exploitation chain. In the Linux kernel-based world, this means that once an attacker takes root on a server, cloud instance, or Android device, they may directly take over the underlying system without needing to breach more external defenses, thanks to the privilege escalation vulnerabilities.
This time, the automated process confirmed and generated usable exploit code and candidate patches for 948 local privilege escalation vulnerabilities, and PoCs for 291 information leakage vulnerabilities, all in a single run. According to public information, the Linux kernel is a core component of global servers, cloud infrastructure, and Android devices. Mass-producing privilege escalation exploits on such a foundational layer is essentially like embedding a complete chain of "invisible bombs" that can be ignited at any moment beneath critical infrastructure. However, the brief explicitly did not disclose whether and how these LPE exploit codes were responsibly handed over to the Linux kernel community, nor did it specify the repair progress of related vulnerabilities, much less any details on whether these vulnerabilities had ever been maliciously exploited. The security community faces a blade whose power has been repeatedly emphasized, yet with unclear control boundaries, leading to an inevitable state of unsettled trust.
From writing code to mining vulnerabilities: OpenAI bets on cybersecurity models
From an external perspective, what has been repeatedly emphasized in this disclosure is not "how many kernel vulnerabilities were uncovered," but the chain referred to as the "automated cybersecurity process" itself—along with its bundling relationship with the "upcoming cybersecurity model." OpenAI has extended its general model capabilities that were originally limited to "writing code" and "modifying code" to security research at the operating system kernel level: a production line starting from generating 2,308 candidate samples progresses through vulnerability discovery, dynamic validation, and feasibility assessment, screening 1,724 triggerable vulnerabilities, before directly producing repair suggestions in the patch generation phase, pairing PoCs and candidate patches for 948 local privilege escalations and 291 information leakage vulnerabilities, before finally handing the results over to human security researchers for final review. For OpenAI, what is showcased is not a "research results release conference" but a set of reusable offensive and defensive infrastructure forecasts.
The stakes behind this combination of methods rewrite the division of labor and rhythm of offense and defense in the security industry: once such cybersecurity models and automated processes are standardized, traditional vulnerability mining and validation performed by small teams over long periods may very likely be compressed into a new paradigm of "model bulk scanning—human sampling review," where the high cost of 1.21 million USD is packaged as the price of a potentially commodity "security computing power service." However, OpenAI has not completely handed the scepter over to machines—while patch generation is included in the process, it deliberately retains manual review as a gate at the end while remaining silent on review ratios, false positive rates, and patch quality standards, indicating that while emphasizing automated efficiency, it is also establishing responsibility boundaries through opaque human-machine collaboration. How the entire industry will align with this new division of labor remains an unresolved open variable.
When AI mines vulnerabilities in bulk, can the rules keep up?
When an automated chain confirms 1,724 Linux kernel vulnerabilities in a single run, generates exploit code for 948 of those local privilege escalation vulnerabilities, and writes PoCs for 291 information leakage vulnerabilities, the production function of security research has been rewritten: what previously required years of accumulated results from small teams can now be produced in bulk at once. Behind the leap in efficiency and scale is a multidimensional dual-use amplification effect—where the same process can help defenders discover and patch weak links in advance, but when abused, can amplify the malicious intent of a very few individuals into thousands of reusable attack paths. The total API cost of 1.21 million USD currently locks this capability within a handful of institutions with the necessary computing power and funding, as the security community monitors the potential impacts of the 948 LPE exploit codes while viewing the cost threshold as the only visible "physical fence" at this stage. However, as these types of cybersecurity models gradually become commoditized, the issue becomes not just who has computing power and who uses the tools first, but: who can access the model, how the calling process is recorded and audited, and the paths along which the generated vulnerabilities and patches should be responsibly disclosed within what time window. Especially since OpenAI has not yet disclosed details of the review mechanism, vulnerability repair progress, and future openness strategy, the industry's consensus on access control, disclosure processes, and user thresholds remains at the discussion draft stage. Looking back at this round of disclosures from the perspective of 2026, both the market and regulatory agencies are still debugging boundaries using a "feeling out" approach: unwilling to let rules bind the hands and feet of the new generation of defensive tools, yet having to admit that when AI begins to mine vulnerabilities in bulk, the existing security and compliance frameworks are no longer sufficient to independently bear the role of risk pricing.
Join our community to discuss and become stronger together!
AiCoin exclusive Hyperliquid benefits: https://app.hyperliquid.xyz/join/AICOIN88
AiCoin exclusive Aster benefits: https://www.asterdex.com/zh-CN/referral/9C50e2
On-chain Telegram community: https://t.me/AiCoinWhaleData
On-chain community: https://www.aicoin.com/link/chat?cid=N6OVMor5g
AiCoin on-chain Twitter: https://x.com/aicoinwhaledata
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
