Break the Chain: Cutting Exploit Paths at AI Speed, Before the Patch
Threat intel always spoke in chains. An actor gets in here, pivots there, lands on the crown jewels. Vulnerability management spoke in rows: one CVE, one asset, one ticket, sorted by CVSS. For years defenders shipped the rows, because chaining was expensive for attackers to pull off at scale. AI removed that cost.
This session is about what that shift means for anyone who owns remediation. We walk a real chain that turned a single foothold in a trusted AI supply-chain platform into admin across multiple production clusters in under 13 hours, assembled by agents that had no map of the environment and brute-forced the path anyway. Every link scored like noise on its own. Together they were a breach.
Drawing on public research from Zafran Labs and others, including the trust_remote_code TOCTOU bypass class behind the FaceHugger disclosures, we break down why single-CVE severity is the wrong unit of decision, how to read exploitability across a real attack path, and where a chain can be broken using compensating controls you already run, before a patch cycle ever starts. Attendees leave with a defender’s method for severing exploit paths at machine speed, and a sharper way to separate the small fraction of exposures that can actually reach them from the flood that can’t.