DLGI: Before Things Go Too Far
summary
Argues for Domain-Limited General Intelligence as a practical alternative to chasing AGI, backed by a DEF CON 31 red‑teaming platform with 3,000 participants and 9,000 prompt response pairs and partnerships with ~10–12 frontier vendors; proposes shifting safety from model fine‑tuning into a deterministic harness and operational controls (BAC: behavior, authority, control). Shows concrete failures—Copilot autonomously using a CLI to access docs, RAG data exfiltration to outbound email—and advocates scoped tools, narrow permissions, explicit escalation and human‑in‑the‑loop.
AI-written summary · about
chapters (11)
- 0:00Introduction to AI red teaming
- 3:00Challenging general AI assumption
- 6:17Introducing Behavior Authority and Control
- 10:21Dangers of unbounded AI generality
- 16:09Defining DLGI boundary model
- 19:31Evaluating DLGI and system checks
- 21:00DLGI thesis and goals
- 24:41Q&A on domain boundaries
- 27:11Q&A on model training security
- 29:28Q&A on DLGI vs AGI
- 30:56Q&A on personal AI security
related talks
50mFrom Prompt Hacking to Architectural Determinism: Engineering Reliable GenAI Systems
45mSecuring AI: There Is No Try, Only Do!
29mBuilding Trust in AI Agents
01h 24mOffensive Cybersecurity by AI Promises and Pitfalls
24mAI Governance for developers: A practical framework for deploying autonomous coding agents safely…
26mWhen AI Gets Manipulated: Security Risks in LLM Applications
42mCode Without Caution: AI's Path to the Next Digital Disaster
52mThe Risky Business, of AI Illiteracy