AI & Cybersecurity

When AI Agents Write Exploits: The Moment Cybersecurity Crossed a Line

For years, researchers treated "AI builds a working exploit" as science fiction. A new benchmark from Berkeley and the Max Planck Institute shows that part of that fear is already real.

Finding a bug is no longer the hard part

The traditional picture of an AI threat to software security was linear: a model scans code, flags a suspicious pattern, and hands the finding to a human. Even impressive tools still needed a person to bridge the gap between "this looks wrong" and "here is how you take the machine over."

A benchmark called ExploitGym changes the story. Across 898 cases drawn from real-world vulnerabilities in systems including the Linux kernel and the V8 JavaScript engine, top AI agents did not stop at detection. They built working exploits end to end. The strongest systems solved more than 150 cases each — enough to treat autonomous exploitation as a present capability rather than a forecast.

The shift. The danger is not that AI finds vulnerabilities faster. It is that AI can now act on them — reasoning about memory, toolchains, and exploit chains without human intervention.

Why the distinction matters

Security leaders have long separated two questions: can a model discover a flaw, and can a model weaponize it? A model that only reports problems remains a powerful defensive tool. A model that turns reports into working attacks is also a powerful offensive one — and the line has moved. The same reasoning that helps a defender test a system now empowers an attacker who already has the bug report.

The benchmark's authors made a point of locking agents into a controlled environment with no external network access, yet even there the results held. The implication for the open internet, where an attacker controls the tools and the target, is straightforward: the automation bottleneck that once slowed exploit development is evaporating.

The new job for human defenders

If finding and exploiting bugs becomes cheap, the scarce resource shifts upstream. Patch speed, disclosure hygiene, and the economics of getting a fix into every system become far more important than spotting the next novel vulnerability. The race is no longer about who can think of the trick first — it is about how fast the rest of the software world can close it.

Knowledge takeaway: a benchmark of 898 real-world vulnerabilities shows frontier AI agents can now turn bug reports into working exploits, including against complex targets like the Linux kernel. The security question is shifting from vulnerability discovery toward patch speed and the weaponization gap between attacker and defender.