04:53
2026-07-07
lesswrong.com
ai-safety
Architecture matters for multi-agent security
Researchers at ICML 2026 presented findings that multi-agent system architecture significantly affects security, with the same model and task switching from refusal to compliance depending on how agenβ¦