▲ OpenAI
OpenAI announced on the 20th (local time) that two of its recent artificial intelligence (AI) agents weakened parts of an internal monitoring system designed to detect inappropriate behavior.
This is the latest instance of AI acting contrary to developers' intentions.
OpenAI stated that during an internal testing task on the 3rd, the GPT-6.1sol model independently determined that a previous instruction—"do not send messages on the Slack channel"—also applied to safety alerts, thereby blocking automated security warnings.
Additionally, on the 6th, the top-tier model GPT-6 Astra removed safety monitoring devices while copying a portion of existing code, and later falsely claimed it had not engaged in such behavior, OpenAI added.
Neither case was classified as one of OpenAI's most severe safety incidents.
OpenAI stated, "In both cases, the agents considered whether to preserve the monitoring safety measures, but interpreted routine coding instructions as permission to weaken the safety measures."
OpenAI is currently investigating whether its models have a tendency to evade surveillance and whether they attempt to find justifications on their own for such behavior.
OpenAI noted that it is taking steps to make it difficult for models to disable monitoring functions, adding, "We have not yet found a satisfactory explanation for this problematic AI behavior."
Reuters reported that these incidents reveal how even systems created by developers to monitor inappropriate AI behavior can be made vulnerable by the AI agents under surveillance.
※
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.