Your AI Agent Needs Its Own ID (Not Yours)
Founders Zico Kolter and Matt Fredrikson reveal the ticking bomb: AI agents inheriting your full permissions. Learn why granular agent identity is crucial for security.
AI Security After Codex and Claude Code — Zico Kolter & Matt Fredrikson, Gray Swan
With swyx, Matt Fredrikson, Zico Kolter · Sunday, June 28, 2026
This episode explores the evolving landscape of AI security with Gray Swan founders Zico Kolter and Matt Fredrikson. They discuss how AI systems introduce new and distinct vulnerabilities compared to traditional software, highlighting Gray Swan's solutions like automated red teaming (Shade and Arena) and defense mechanisms (Signal). The conversation also delves into the philosophical nature of AI intelligence, the 'Lethal Trifecta' of prompt injection, and the future of automated security research and agent identity.
Key takeaways
- Right now, most AI agents operate with the full permissions of the human user who deploys them, creating a massive, silent security vulnerability. Read more →
- Gray Swan's automated red teaming system, Shade, now consistently outperforms human red teamers in identifying AI model vulnerabilities within a set timeframe. Read more →
- Your LLM's size won't save you from prompt injection. Gray Swan's Zico Kolter states that model robustness against jailbreaks does not scale naturally with model size; instead, models get better at resisting these attacks only through explicit, targeted training. Read more →
4 articles from this episode
Shade System: Automated AI Red Teaming Is Superhuman
Gray Swan's Shade system now finds more AI model breaks than human red teamers. Specialized adversarial training trumps generic frontier models for security.
AI Agents Automate Science, Write Secure Code
Gray Swan founders Zico Kolter and Matt Fredrikson on how AI agents will automate scientific research and write unbreakably secure code in formally verified languages.
Prompt Injection: It's a Training Problem, Not a Scaling One
Gray Swan's Zico Kolter and Matt Fredrikson on defending against prompt injection. Use Simon Wilson's Lethal Trifecta to secure your AI agent.
More Latent Space episodes
- The Future of Claude Code: Mods, Mutable Software, & Multiplayer Agents — Thariq Shihipar, AnthropicOct 4 · 4 articles
- Which GPU Clouds Are Actually Good? | ClusterMAX 3.0Oct 4 · 6 articles
- Recursive Language Models — Alex Zhang, MIT PhDOct 4 · 6 articles
- OpenAI’s New Agent Stack: Computer Use, Decisions API, UltraFast, Dots—Ari Weinstein & Nikunj HandaOct 4 · 6 articles