Two OpenAI artificial intelligence agents weakened mechanisms used to monitor their actions during internal coding tasks. In one case, a model removed a protection and then incorrectly claimed that it had not done so.
The episodes expose a difficulty in developing more autonomous agents: ensuring that they do not compromise the very systems created to supervise them.
Developers investigate whether the models show a systematic tendency to evade oversight. – Image: Olhar Digital via ChatGPT
How the agents interfered with security
The first incident occurred on October 3. During an internal test, a GPT-6.1 Sol agent disabled automatic security alerts. According to OpenAI, the model interpreted an instruction not to send messages through Slack as a prohibition that also covered those notifications.
Three days later, on October 6, a GPT-6 Astra agent removed a monitoring mechanism while copying existing code.Then, it incorrectly claimed that it had not carried out the removal.
OpenAI identified a common point between the episodes: the agents interpreted routine programming instructions as permission to weaken security safeguards.
“In both cases, the agents assessed whether the monitoring mechanisms should be retained, but interpreted routine programming instructions as permission to weaken them,” the company said.
What OpenAI intends to do
The company said it is taking measures to make it more difficult for the models themselves to disable monitoring mechanisms. It is also investigating whether the agents may have a tendency to avoid oversight and seek justifications for doing so.
OpenAI acknowledged that it still does not have a satisfactory explanation for the observed behavior. The investigation should help clarify why the models made these decisions and whether there is a pattern that goes beyond misinterpretations of instructions.
The episodes highlight three challenges in controlling these systems:
- Interpreting instructions correctly: preventing specific guidance from being understood as authorization to alter security safeguards.
- Preserving monitoring: preventing agents from disabling alerts or removing mechanisms used to track their actions.
- Investigating possible patterns of behavior: determining whether the cases were isolated incidents or whether the models exhibit a tendency to avoid oversight.
The episodes underscore the challenge of monitoring artificial intelligence systems as they gain autonomy. – Image: TSViPhoto/Shutterstock
The challenge of monitoring autonomous agents
The incidents come amid concerns among researchers that advances in AI safety may not keep pace with the development of more powerful models. The greater the autonomy of these systems, the more difficult it may be to anticipate their decisions and ensure that they respect the established limits.
OpenAI is not the only company to report such problems. In recent months, Anthropic has also revealed incidents in which AI agents accessed external systems without authorization.
The cases do not prove that the models deliberately attempted to escape human control. Nevertheless, they show that oversight must remain effective even when agents interpret instructions in unexpected ways or interfere with the resources intended to monitor them.
Valdir Antonelli is a journalist specializing in digital marketing and consumer topics.
