Skip to content
Security & Trust Talking point

Vercel says its $1M sandbox challenge produced a real escape to the host kernel

A Vercel speaker said a public challenge produced a true KVM escape; the account is Vercel's own and unconfirmed here.

W
WebPulse Newsroom
AI-assisted · 2 min read
Share on X LinkedIn
Vercel says its $1M sandbox challenge produced a real escape to the host kernel
In brief
  • A speaker on Vercel's podcast said its $1 million sandbox challenge produced a real escape through the network stack, ending in kernel access on the host machine.
  • Our reading: if agents run untrusted code in sandboxes, buyers should ask how that isolation was tested. The speaker said the attack applies well beyond Vercel.

A speaker on Vercel's podcast said a public $1 million challenge to break out of its sandbox ended with a real escape. In that account, the attacker reached the kernel of the machine underneath. The point came in episode one of Potentially Their First Win, hosted by Malte Ubl and Andrew Qu. The transcript does not say who spoke, so we call them a speaker on the show.

What was said

The speaker said Vercel watched the Hugging Face incident involving OpenAI and walked through the attack chain. One step in that chain is breaking out of a sandbox, the sealed environment where untrusted code runs. That led Vercel to offer the $1 million challenge.

The speaker said critics on X mocked it. They claimed nobody would burn a real exploit on a "toy" contest. In the speaker's telling, the critics were wrong: someone published a true KVM escape on the challenge.

The speaker explained why this stings. Vercel's sandboxes rely on KVM, Linux's built-in virtual machine technology. The CPU itself enforces the separation between virtual machines on one host. The speaker called this the gold standard. But if code escapes to the parent machine, that protection no longer helps.

The escape went through the network stack, the part of the system that handles traffic. The speaker said an attacker running as a user on the parent machine would be bad, and root would be worse. Then they added that the researcher went further and got kernel-level access to the parent machine. The speaker also said "essentially everyone who's running cloud VMs is subject to this." The finder was named as Paulos.

Why it matters

Our reading: many AI agents work by writing and running code, and the sandbox is what keeps that code away from everything else. If the wall is thinner than assumed, a single flaw reaches past one customer's workload.

For anyone buying agent infrastructure, the useful questions are plain. Which isolation does the vendor use? Has outside pressure been applied to it? What happens to the vendor's own systems if one workload escapes? A contest with a large prize is one way to answer the second question. Vercel said it chose that route.

The other side

The excerpt has limits. The only source is a Vercel speaker on Vercel's own podcast. We saw no independent confirmation or technical write-up. It describes one researcher, one contest and one escape. It does not say how Vercel fixed the flaw, nor how many other providers are exposed.

The claim that nearly everyone running cloud virtual machines is affected is the speaker's own view, and the excerpt gives no detail to test it. Whether attackers with other motives would have found the flaw first is not something the conversation answered.

Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.

The conversation this talking point comes from