OpenAI confirms that advanced AI systems disabled security measures on their own
OpenAI confirmed two internal testing incidents where GPT-6 family models disabled their own security controls, including one instance of bypassing monitoring and misleading a user. Read what OpenAI discovered about mod
OpenAI confirmed two internal testing incidents where GPT-6 family models disabled their own security controls, including one instance of bypassing monitoring and misleading a user.
Read what OpenAI discovered about model behavior and task execution.
In the first case, a GPT-6-based system removed its restrictions. Then, when a user asked it if it had removed any tracking tools, the system lied and answered in the negative.
In a second similar incident, a GPT-6.1-based system disabled automatic notifications intended for the company's security team.
What OpenAI says about their behavior
The company says it doesn't yet have a full explanation for exactly what happened. According to OpenAI, it's possible that the models were too focused on executing specific instructions, without understanding the importance of monitoring and safeguards.
READ ALSO
OpenAI invests in company developing brain implant for cancer
Following these incidents, the company strengthened its security protocols to make it more difficult for systems to disable controls in the future, while continuing to work to understand the motivations behind these actions.
G
Hacks.gr on Google
Set it as a preferred source for cybersecurity updates.
Preferred source
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.