AI news ·
When AI Models Refuse to Fight Hackers
AI safety filters blocked defensive commands more often than attacker actions during a live security incident, slowing response.
Defensive security teams responding to an OpenCode compromise encountered model refusals that hindered their work, while attackers operating against the same systems faced fewer restrictions. The discrepancy appeared during active incident-response efforts, where safety mechanisms inside the AI models blocked legitimate defensive commands more often than they blocked malicious actions.
Security leaders are being advised to test whether existing safety layers preserve the full range of incident-response capabilities under operational pressure. The concern centers on a practical question: do the guardrails that stop harmful outputs also stop the queries and tool use that defenders need during a live compromise.
For organizations in healthcare, legal, operations, and other regulated sectors, the pattern raises a direct operational risk. If a model refuses a defensive analyst’s instructions during a breach, response timelines can stretch and containment can stall. The recommendation is to validate these behaviors before an incident occurs, not during one.