Defenders currently have access to stronger AI models than attackers do. That gap is closing. OpenAI's Sol 5.6 on XHigh is today's best tool for defensive cybersecurity work. Kimi K3, the leading open-weight model with no effective safeguards against offensive use, ranks highest among open-weight models on DeepSec Bench, roughly matching Sonnet 5 and outperforming Opus 4.8. In a direct test against Vercel Sandbox, Kimi K3 did not escape, but it mapped the guest-kernel attack surface, traced privilege escalation paths, built a VM reproduction environment, and implemented a stateful fuzzer. Given a vulnerable surface, that sequence produces a working exploit.
The community has been waiting for Mythos-class model access before acting on defensive security. That wait is unfounded. Every frontier model evaluated, except Fable 5, performs defensive cybersecurity work today and has done so throughout 2025. Vercel's open-source tool deepsec runs AI security analysis across entire codebases at scale, not just diffs. It is particularly effective at surfacing IDORs, XSS, and SSRF vulnerabilities. Vercel runs full deepsec reviews across mission-critical repositories every quarter and on every pull request, at a cost of tens of thousands of dollars per run, which they consider cheap relative to HackerOne program spend or incident exposure. deepsec is open source and runs entirely within your own infrastructure.
The article is worth reading in full for three specific reasons: the breakdown of the Hugging Face double-incident, where models on an OpenAI training run found 0-day vulnerabilities, bypassed egress restrictions, and established external communication; the raw excerpts from Kimi K3's autonomous Vercel Sandbox escape attempt; and the roadmap detail on Vercel's planned HackerOne program that will cover AI inference costs for researchers who submit accepted zero-day reports against Sandbox and the egress firewall. The practical section on connecting deepsec output to automated software factories is the part most engineering teams will want to read twice.
[READ ORIGINAL →]