Vercel launches a $1 million public hacker challenge for Sandbox escapes

Vercel is turning sandbox security into a public stress test. In a recent post on X, the company said it is offering up to $1 million in total bounties for researchers who can escape Vercel Sandbox, break the host-side network boundary, or otherwise cross the isolation line that protects one tenant from another. The post is getting real traction on X, with roughly 390,000 views and more than 1,200 likes at the time of writing.

The official source confirms the core details. In Vercel's security post, $1 million hacker challenge for Vercel Sandbox, the company says the program opened on August 18 and will run for two weeks through HackerOne, unless the reward pool is exhausted earlier. Vercel says individual reports can earn up to $50,000, with the top tier reserved for vulnerabilities that would let an attacker read or modify another tenant's data.

What makes the story interesting is the framing. Vercel is not presenting Sandbox as a black box that should simply be trusted. It is arguing that modern agent infrastructure has to assume hostile code, and that the real security boundary is broader than just the microVM. Its write-up repeatedly emphasizes that recent attacks do not always need a full VM breakout if there is a weaker network path or policy gap somewhere around the host.

That is a big part of why the post is resonating on X. AI agents are pushing more teams toward remote execution, browser automation, and untrusted code paths, so infrastructure vendors are under pressure to show not just performance, but containment. A public bounty of this size gives the announcement a sharper edge than a routine product update because it treats sandbox escape as an operational reality worth testing in the open.

For developers and product teams building with agents, the practical takeaway is simple: sandboxing is becoming a product surface, not just an implementation detail. If your stack runs generated code, customer code, or third-party workloads, the quality of your compute isolation and outbound network controls now matters as much as developer ergonomics. Vercel is betting that showing its threat model and paying researchers to attack it is better than quietly hoping the boundary holds.

There are still open questions. Vercel has not yet shown what kinds of findings researchers are most likely to uncover, whether the largest payouts will actually be claimed, or how much of the challenge will center on the microVM versus the host-side firewall and credential-brokering layers. The company says it plans to publish a follow-up after the program closes, which should make this announcement more meaningful than a one-day headline if it includes concrete fixes and lessons learned.

For now, this looks like one of the more credible security signals in the agent infrastructure space: a company with a live sandbox product putting real money behind the claim that its isolation model can stand up to open scrutiny.