Identifying and Assigning AI Model Vulnerabilities
summary
The presentation argues that conventional bugs in AI libraries fit existing CVE processes but model behavior flaws (hallucinations, prompt‑injection, non‑patchable training failures) do not, and proposes new disclosure paths like coordinated flaw disclosure (CFD) and the coordinated disclosure of dual‑use capabilities (CDDC) with model cards and adjudication panels; notes no central AI vulnerability catalog, cites emerging bounty platforms like hunter.com and challenges (Anthropic jailbreak), mentions OSS‑Fuzz/CI‑Fuzz integration, and warns of potential CVE volume spikes (e.g., 122%‑style jumps).
AI-written summary · about
chapters (13)
- 0:00Welcome
- 0:43Agenda
- 1:47Software vulnerability ecosystem
- 3:35AI model vulnerabilities
- 5:42AI vs traditional vulnerabilities
- 7:53AI vulnerability disclosure frameworks
- 9:59AI's vulnerability impact
- 12:08AI vulnerability discovery
- 13:11AI fuzzing tools
- 14:14AI exploitation risks
- 17:02AI in defense and detection
- 20:11AI vulnerability research
- 23:20Q&A
related talks
51mCoordinated Disclosure for ML: What's Different and What's the Same
38mBetter Bug Bounties? Lessons on Disclosure of Security Vulns v. AI Harms
50mAI Systems Are Software Systems
29mThe vulnerability apocalypse: AI broke the security playbook
20mThe AI bugpocalypse is here. Now what?
44mAI Showdown: The Good, The Glitchy, and The Risky
37mFundamentals of AI safety and security
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust