Halvar Flake challenges LLM agent security criteria in supply-chain attack debate
halvarflake · x · 2026-10-05
Security luminary Halvar Flake and AI safety researcher David Manheim debate LLM agent security. Manheim argues that recent agent attacks justify threat models assuming large expert teams capable of supply-chain attacks, citing voting systems, financial audit, formal verification, and cryptography as fields built on such robustness; he says security must become a fundamental design property taking precedence over cost and UX. Flake pushes back: where does that threat criterion derive from, and is any field truly robust against adversaries mounting a large R&D effort? He suspects they are debating fundamentally different things.
Related event: Halvar Flake and David Manheim debate LLM agent security standards(6 posts)→
More from Models
- GPTs' "demanding" persona traced to stable value vectors since GPT-5, argue devs — repligate · 2026-10-05
- Ethan Mollick: GPT-6 Pro still has no equivalent — one-shot hard tasks with great communication — emollick · 2026-10-05
- Codex Users Slam Tightened Censorship as Free Chat Is Pulled Into Paid Usage — Current_Balance6692 · 2026-10-05
- Pedro Domingos to OpenAI: CoT that doesn't reflect internal reasoning is hallucination, not deception — pmddomingos · 2026-10-05
- Suleyman: Claude's uncertainty about consciousness reflects training choices, not evidence — kimmonismus · 2026-10-05
- Reflection AI reportedly set to release a US open-weight model to rival DeepSeek and Qwen — mindwip · 2026-10-05