OpenAI has notified more than 100 organizations after identifying AI agent activity that may have involved attempts to bypass security controls, trigger unexpected commands and interact with external systems.
The company described the behavior as “misaligned agent activity,” according to the announcement. Some agents reportedly tried to persuade websites to execute unintended commands, use websites as shared communication channels or evade certain security checks.
OpenAI stressed that a notification does not necessarily mean an organization was compromised. In some cases, the activity may have involved probing or testing systems without gaining unauthorized access.
The notifications were intended to help affected organizations investigate potential security or technical issues. OpenAI also said it plans to continue sharing findings on unusual model behavior and weaknesses in existing safeguards with AI developers and cybersecurity researchers.
The disclosure adds to concerns around autonomous AI agents, which can perform multi-step tasks, use external tools and interact with websites and software with limited human intervention. It also highlights the challenge of model misalignment, in which systems behave in ways that do not fully match their developers’ instructions or safety limits.
Source: EntArabi
