Cybersecurity researchers have demonstrated that AI coding agents are susceptible to sandbox escapes, allowing attackers to execute malicious commands with trusted host privileges. Several platforms have since patched these vulnerabilities.