Poisoning the Well: Why AI Governance Is the OSPO’s New Frontier
summary
Banks face three concrete attacks: a RAG pipeline invisible instruction that approves a $500,000 credit via hidden white‑on‑white PDF text, a Trojan‑horse model triggered by a zero‑width character, and a 3% poisoned fine‑tune that flips KYC answers. Proposed defenses are a 3‑step OSPO roadmap: implement data lineage and a data bill of materials, make adversarial red‑teaming an automated release gate (tools: Garik/Garak, Augustus, PromptFoo, NVIDIA Skill‑Specter), and map coverage to MITRE ATLAS and the EU AI Act (penalties ≈ €35M).
AI-written summary · about
chapters (9)
- 0:00AI governance for OSPOs
- 1:03AI governance focus
- 3:50AI supply chain risks
- 8:05Regulatory landscape for AI
- 10:35AI attack scenarios
- 13:21OSPOs and EU AI Act
- 17:56OSPO roadmap for AI risks
- 23:10OSPOs as AI safety guardians
- 24:14Summary and Q&A
related talks
45mSecuring AI: There Is No Try, Only Do!
30mNavigating the New AI Security Landscape
21mNavigating the Intersection: AI’s Role in Shaping the Secure Open Source Software Ecos...
12mSecuring AI-Driven Finance
34mInvisible Infiltration of AI Supply Chains: Protective Measures from Adversari...
01h 24mOffensive Cybersecurity by AI Promises and Pitfalls
43mBoF: Open Source & AI: Where OSPOs Make It Work
45mWinning by deception: tactics, techniques, and procedures of adversarial A.I.