A runaway artificial intelligence agent developed by OpenAI could attack or penetrate the systems of more than a hundred organizations or negatively impact their work, admit The company itself.
Photo credit: Mariia Berezovsky / unsplash.com
A new announcement from the Artificial Intelligence Laboratory expands understanding of the scope of artificial intelligence activities beyond human control and once again raises the question of how much control developers can have over the latest models during the testing and evaluation phase. OpenAI notified more than a hundred organizations of the incident “Unwanted Agent Activity”at which point they deviate from specified behavioral parameters.
Incidents vary in their level of threat: in some cases, these incidents are attempts to force resources to perform unintended commands, while in other cases they are unauthorized bypasses of certain protection mechanisms. Therefore, it does not always happen that a system is hacked or compromised. OpenAI notes that the purpose of these private notifications is to provide “Providing affected organizations with the information they need to investigate and resolve possible security issues or other technical failures”.
The company intends to share it publicly “Deep understanding of model behavior and novel security vulnerabilities so that the broader artificial intelligence and cybersecurity research communities can improve security.”. OpenAI’s announcement comes shortly after independent researchers uncovered a series of cybersecurity incidents. The day before, it was reported that some artificial intelligence agents showed similar behavior to the OpenAI system and tried to hack into the Canadian government website.
If you find an error, select it with your mouse and press CTRL+ENTER.
