Top AI coding agents can be easy victims to sandbox escapes, showing they aren't as secure as they claim to be

A robot hand touching a locked digital shield blocking a human from accessing data.

• Pillar researchers demonstrated sandbox escapes in AI coding agents
• Exploits let attacker‑written configs run with trusted host privileges
• Agentic security needs its own threat model, researchers claim

AI coding agents can be tricked into turning on their operators and assisting attackers in compromising the underlying systems, experts have warned.