AI pipeline generates 948 working Linux exploits, revealing scale of automated hacking
An AI‑driven system has automatically produced 948 working Linux exploits and reproduced 1,724 vulnerabilities in 30 hours, at an estimated API cost of $1.2 million, underscoring how automation could change the economics of software attacks.
A recent large‑scale test of AI agents on Linux systems offers a concrete look at how automation could change the balance between attackers and defenders in cyberspace.
Researchers report that an AI‑orchestrated pipeline, running for 30 hours, was able to reproduce 1,724 vulnerabilities on Linux and generate 948 working local‑privilege‑escalation exploits. These are programs that let an attacker turn limited user access into full root control. The system also produced 291 proof‑of‑concept information‑leak attacks. The estimated cost of the AI API calls that powered the run was about $1.2 million.
The project focused on Linux, the open‑source operating system that underpins much of the world’s server, cloud, and embedded infrastructure. By chaining together automated tasks such as gathering vulnerability data, writing and testing exploit code, and iterating on failed attempts, the pipeline showed that much of the exploit‑development process can be handed to machines. Humans still had to supervise, but their role shifted from hand‑crafting individual exploits to steering a high‑volume system.
For defenders, from corporate security teams to government agencies, that shift has immediate implications. Security staff already struggle to patch vulnerabilities quickly, and many organisations lack enough skilled personnel. If well‑resourced attackers can generate hundreds of working exploits against older or poorly maintained systems, defenders may find themselves facing more intrusion attempts than human teams can realistically manage.
On the offensive side, states and criminal groups stand to gain new capabilities. Intelligence agencies that once depended heavily on specialist exploit developers can use AI to expand their arsenals, particularly for less complex but still damaging local attacks. Ransomware operators or other financially motivated hackers may not need deep knowledge of operating system internals if they can access or copy an AI pipeline that performs that work for them. The main constraints become budget, computing power, and access to advanced models.
Strategically, this kind of automation puts pressure on existing norms around responsible disclosure and the stockpiling of vulnerabilities. Governments that hold unpatched bugs for espionage or offensive use must now assume that AI‑assisted adversaries can rediscover some of those vulnerabilities at scale. Software vendors face attackers who can systematically mine weaknesses across entire product lines rather than focusing on a handful of high‑profile flaws.
The experiment also highlights a dual‑use dilemma. The same kind of AI system that can help internal security teams test their own networks and strengthen defences can help malicious actors probe for ways in, depending on who controls it and how it is used.
The main takeaway is straightforward: when AI systems can mint working exploits by the hundred, software bugs cease to be a niche problem for small expert teams and start to look like an industrial resource.
The next steps to watch will likely come from several directions. Policymakers may explore new limits or reporting requirements around the use of powerful AI models for offensive security work, though enforcement across borders will be difficult. Major cloud providers and AI firms will face questions about whether and how to monitor large‑scale exploit‑generation workloads on their platforms. Inside security operations centres, teams are already under pressure to adapt similar automation for defence, from prioritising patches to spotting exploit patterns in logs before attackers can do serious damage.
Sources
- OSINT