Exploiting Vulnerabilities at scale with AI

Co-Founder & CPO (Quantro Security, Inc.)

For decades the economics of offensive security favored defenders. Turning a CVE into a working exploit took a skilled researcher days to weeks, and that cost quietly limited how much of the vulnerability backlog ever got weaponized. It doesn't limit much anymore.

‍

We built an LLM-driven exploitation pipeline, ran it against 3,029 real CVEs, and measured what it costs an autonomous agent to reach a verified working exploit. Verified means we reproduced it, not that the model reported success. The median exploit took about 13 minutes and roughly $3 of inference. The human benchmark for the same work is measured in weeks. We drove 72% of analyzed CVEs to a verified exploit, and a substantial share of those have never appeared on CISA's KEV catalog. "Not yet exploited in the wild" is no longer a useful proxy for "expensive to exploit."

‍

The talk treats exploitation as a unit economics problem. We'll cover how we measured it, where the measurement breaks down, and what the CVEs that stayed Unsolved had in common. Then the distributions themselves: cost per exploit and time per exploit across the corpus. The rest is about what defenders do with that. Patch prioritization, KEV-weighted risk scoring, and attacker ROI all assume exploitation is scarce, and none of them hold up well once a working exploit costs about as much as a coffee. We finish on the other side of the ledger, with AI-generated remediation measured the same way, in minutes and dollars.

‍