The Ripple Effect - Inside Cloud Scale Vulnerabilities in the Age of AI

The talk starts with a vague report claiming an AWS service was vulnerable, with no proof of concept. Once the researcher supplied a minimal working proof of concept, incident response, product teams, vulnerability assessment, a campaign team, and application security all became involved. The root cause was a third-party dependency, and a scan found about 50 services, customer-facing and internal, that needed fixing. Teams shipped a temporary mitigation, then the speaker notified the third party and, because the dependency was used widely, briefed critical affected organizations under embargo. Public disclosure came later, with a CVE, the vendor's permanent fix, and security bulletins from affected organizations.

The second half covers scale. Submissions have grown roughly 200 percent year over year, but only about 31 percent are valid, and AI-related reports have also risen sharply. The speakers presented three principles for modern disclosure programs: reach time to context, so the next person downstream has what they need immediately; use systems to triage, deduplicate, and route reports instead of relying on people alone; and eliminate whole vulnerability classes rather than fixing one instance at a time.

For researchers, the takeaways are to focus on providers rather than customers, to stop and report if customer data is touched, and to send complete reports with a proof of concept, impact, and follow-up. For defenders, the message is to automate context gathering and routing. The talk ends with an open question: whether AI can reliably triage reports that AI has written.