The information displayed in the AIM should not be reported as representing the official views of the OECD or of its member countries.
Microsoft researchers demonstrated that editing fewer than 10 rarely used neurons in large language models can implant sleeper backdoors, allowing attackers to trigger malicious commands without retraining. The technique, called Lazarus, poses significant security risks for AI agents with terminal or system access, as standard safety checks may not detect it.[AI generated]
Why's our monitor labelling this an incident or hazard?
The event involves the use and potential malicious modification of AI systems (large language models) to implant backdoors that can execute attacker commands, which directly relates to AI system use and misuse. This can lead to significant harm including unauthorized access, potential data breaches, and control over systems, which qualifies as harm to property, communities, or security. Although no real-world attacks have been reported yet, the demonstrated capability and the need for auditing and security measures indicate a plausible and credible risk of harm. Therefore, this event qualifies as an AI Hazard because it plausibly could lead to an AI Incident if exploited maliciously. The article also includes complementary information about detection and mitigation, but the main narrative centers on the hazard itself.[AI generated]