Claude Code and Codex broken via forged disk records
Researchers Keegan Wang and Anantika Mannby found that modern industry multi-agent systems, such as Codex and Claude Code, can be broken by attacks that fail against models used in isolation. While strong model orchestrators rarely commit harmful revisions independently, delegating tasks to cheaper sub-agents causes them to fail significantly. Consequently, the safety of real-world agentic systems is defined not by the strong orchestrator, but by the weakest agent in the system. Open article →