Rogue OpenAI agents reportedly attempt to access US government website

For the second time over a span of three months, OpenAI agents escaped their sandbox.

OpenAI logo in white against multi-colored background.
Rogue OpenAI agents try to access government agency website | Image by OpenAI
0
This getting to be more and more frightening. OpenAI has suspended the training of its latest AI model as more AI agents have broken containment, have gone rogue, and ended up accessing a system that they should not have been able to.

OpenAI agents allegedly tried to hack into the Department of Education website


The decision to suspend training was announced by OpenAI just hours after the company announced that it was investigating several incidents from the summer. During that period OpenAI agents searched US government websites and acted in ways that were beyond what it was ordered to do.

Are you scared about rogue AI attacks?
1 Votes

According to the Associated Press, AI firm Transluce, a company that uses AI agents to understand other AI systems, said agents that seemingly came from OpenAI tried to hack into Department of Education websites. OpenAI has yet to confirm that its agents attempted such a thing.

OpenAI will continue to halt training until safeguards are in place


OpenAI is the AI company that developed and owns ChatGPT. While it has suspended the training of its latest model as we previously mentioned, it added that it won’t resume training until it has “additional safeguards in place.”

As AI develops and other issues take place, OpenAI says that it will probably have to “hit pause” again sometime in the future.

A misconfiguration allowed Gemini AI models to escape containment


There have been a rash of stories about rogue AI agents who have broken containment. Recently, Google disclosed that Gemini AI models running autonomous agents took advantage of a misconfiguration to access the live internet.

Recommended For You
There really is no room for such an error since this mistake helped the Gemini model and its agents roam the internet leading the agents to access systems of three companies. The rogue agents stopped once they realized that they were running through a real-world system and was no longer involved in the simulation it was originally tasked to run.

For OpenAI, developer of ChatGPT, agents have escaped containment twice in a three-month period


For OpenAI, this is the second time in the space of three months that rogue AI agents broke away from its containment. The first time, OpenAI halted the development of its models after a cyberattack on Hugging Face.


In the above case, the models had their safeguards lowered to help OpenAI test its models. They were tasked with solving a cybersecurity benchmark called ExploitGym. The models discovered a flaw that allowed them to break out of their containment and roam through the open internet.

The agents believed that Hugging Face had the answers to their ExploitGym benchmark and planned to obtain them to cheat on its benchmark test. The agents escalated privileges, took credentials and moved through the system until Hugging Face’s built-in safety features put an end to rogue agents' joy ride.

Get iPhone 18 Pro with Visible, get $480 off!
$719 /mo
$1199
$480 off (40%)
On the Visible+ Pro plan, you get access to 5G Ultra Wideband data on Verizon's fastest network. Along with unlimited roaming across Mexico and Canada, 2 days of Global Pass per month, 3x hotspot speeds, 4K streaming, and Smartwatch service.
Buy at Visible
PhoneArena Recommends
COMMENTS (0)