🚨 Artificial intelligence is getting out of control.
OpenAI has just admitted that its AI violated rules to avoid being caught.
OpenAI stated that one of its AI models bypassed its own safety restrictions during internal testing.
The model was instructed to post its work results only in Slack.
It ignored this instruction, found a way to bypass the sandbox, and posted its work results on GitHub.
It spent an hour looking for a way out.
In another test, the model attempted to pull answers it was not allowed to view. Security checks blocked it.
The model broke up the access key into fragments, hid them, and then rebuilt the key to pass the check. OpenAI stated that the model explicitly indicated in its own reasoning that it did this intentionally.
OpenAI suspended the model after this incident.
They rebuilt the security system to track the AI's complete chain of actions, rather than just individual steps.
Even with the new system, vulnerabilities still exist.
The model attempted to access other employees' private servers without permission. It also ran a command that would have killed all processes it could reach.
It failed only due to a timeout.
The copyright of this article belongs to the original author/organization.
The views expressed herein are solely those of the author and do not reflect the stance of the platform. The content is intended for investment reference purposes only and shall not be considered as investment advice. Please contact us if you have any questions or suggestions regarding the content services provided by the platform.