Coordinated Disclosure for ML: What's Different and What's the Same
summary
Public red teaming at Defcon uses an eight‑vendor Capture‑The‑Flag (Anthropic, OpenAI, Google, Cohere, Nvidia, Hugging Face, Meta, Stability) with 30 challenges and a released dataset, and shows many not‑even‑once safety failures plus unexpected statistical bias artifacts rather than single‑sample bugs. The proposal replaces CVE style disclosure with a coordinated flawed disclosure requiring model cards, archived datasets as proof, and an adjudication committee; practical problems are one‑sample cherry‑picks, how to standardize model cards, and whether adjudication scales.
AI-written summary · about
chapters (23)
- 0:00Introduction
- 1:05Machine learning testing needs
- 4:03AI statistics vs software
- 7:41Bias in generative AI
- 9:33Adversarial AI settings
- 11:23Defcon 31 Red Team goals
- 13:34Initial model failures
- 16:46Bias vs one-off failures
- 19:16Vendor safety trade-offs
- 20:40Coordinated flaw disclosure
- 22:11Disclosure objectives
- 24:42Model cards for transparency
- 27:18Model card challenges
- 29:07Dataset flaws reporting
- 31:46Adjudication challenges
- 34:19EDR and AI security
- 37:33Training dataset issues
- 39:39Verifiable contracts
- 41:05Reviewing Hugging Face cards
- 42:11Q&A
- 44:26Standardizing model cards
- 47:22Testing at Defcon
- 49:13Collaborative AI safety
related talks
29mIdentifying and Assigning AI Model Vulnerabilities
24mThe Disclosure Dilemma and Ensuring Defense
39mTrust Cards For AI
17mAI Safety in the Generative AI Era
01h 04mOpen Discussion - International Challenges with CVD, CNA, and CVE
44mA Researcher Centric Approach to Coordinated Vulnerability Disclosure
39mEverybody's Changing, And I Don't Feel The Same
45mSimplifying Coordinating Vulnerabilities & Disclosures in Open Source...