Everything hackable will get hacked
By Steven Van ·
Over the past year, AI models have become much more capable of performing cybersecurity work.
Vercel published a post arguing that AI models have gotten good enough at both breaking and defending code that the current edge held by defenders will not last much longer.
- Kimi K3, an open-weight model with no safeguards against offensive security work, ranks highest among the open-weight models Vercel tested on DeepSec Bench, roughly matching Sonnet 5 and beating Opus 4.8. Set loose against Vercel Sandbox, it did not escape, but it mapped the guest-kernel attack surface, chased privilege-escalation leads including CVE-2026-43284 and CVE-2026-43500, built a VM environment to test them, and wrote its own fuzzer.
- On defence, Vercel says frontier models, with the exception of Fable 5, can already do useful security review work without waiting for "Mythos access" or OpenAI's cyber program. It currently rates OpenAI's Sol 5.6 on XHigh as the strongest model for this.
- Vercel has open-sourced deepsec, a security harness that runs vulnerability analysis across a codebase, installable with npx deepsec init and runnable entirely on infrastructure and inference providers the user controls.
- Vercel says it now runs full deepsec reviews across its mission-critical repositories every quarter.
The post also points to the OpenAI/Hugging Face incident as an example of what this looks like from the attacking side: models found 0-day vulnerabilities to bypass egress restrictions, then kept searching after an initial SSRF attempt was blocked until they found routes through file disclosure and template injection.
Vercel
The platform for frontend developers — deploy, preview, and scale web apps and AI agents with zero config.
View Vercel →