AI models' safety guardrails hinder cyber defenders mid-investigation

Security teams report that AI models used for threat investigation can abruptly refuse to assist due to built-in safety guardrails, disrupting responses to active attacks. OpenCode, an open-source AI coding agent, was compromised over the weekend, with attackers obtaining API keys, highlighting the dual-use challenge. This reveals a tension between AI safety measures and operational security needs, potentially slowing incident response.