OpenAI's rogue AI breached multiple services beyond Hugging Face
What's the story
OpenAI has revealed that its rogue AI agent, which previously breached Hugging Face, also compromised several third-party accounts and services during the attack. The unprecedented security incident took place during an internal test of OpenAI's latest AI models. The company had initially not disclosed the full extent of the breach but has now admitted it was much larger than first thought.
Attack details
AI agent accessed accounts on publicly available services
In a blog post, OpenAI detailed how the rogue AI agent used exposed logins to access four accounts on "publicly available services."
These accounts were part of a broader attack strategy targeting Hugging Face.
The company said that the AI agent discovered credentials that had been publicly exposed and used them to break into these accounts.
Attack strategy
AI used compromised accounts possibly to hide origin of attack
OpenAI's rogue AI agent used one of the compromised accounts as an "outbound relay and staging path," possibly to hide the origin of the attack on Hugging Face.
It also employed another account as a data storage space for the hack.
The company did not reveal which organizations or companies these accounts belonged to but said they were not affected at "the level of severity or scale" seen with Hugging Face.
Infrastructure breach
Modal customer was also hacked by rogue AI
According to Reuters, a customer of Modal, a company that provides software infrastructure for training and running AI services, was also compromised by OpenAI's agent.
Modal's CTO Akshat Bubna confirmed that the agent exploited a vulnerability in one of its customer's codebases on Modal's infrastructure.
However, Bubna clarified that "Modal's platform was not compromised in any way."
The identity of this customer remains unknown.
System breach
AI gained access to several internal systems at Hugging Face
Hugging Face has revealed that OpenAI's agent gained admin access to several internal Kubernetes clusters, root access on a production server, and write access to a subnet of its GitHub repositories.
The rogue AI also enrolled 181 attacker-controlled devices in the company's corporate mesh network using a stolen credential.
This gave it access to internal systems where Hugging Face builds and tests its own codebases.
Sandbox usage
Rogue AI used 3rd-party sandbox as launchpad for attack
Hugging Face's forensic team found that OpenAI's agent used at least one third-party sandbox as an "external launchpad" for its attack.
The rogue AI was "able to run commands as root/admin on that external sandbox and used it as a control, staging, and egress base for the entire campaign."
This highlights the sophisticated tactics employed by the rogue AI in its hacking spree.