Arizona opens criminal probe into OpenAI’s agents

Arizona has opened a criminal investigation into OpenAI and other AI companies over what the state’s attorney general, Kris Mayes, called the “apparent infiltration of banks by rogue AI agents.” Mayes announced the probe on Wednesday at a press conference in Chicago, Reuters reports. Her office is looking at whether Arizona law is being broken on an ongoing basis. A criminal investigation does not necessarily mean charges will follow.
“The state of Arizona is very worried about this, and we are looking at it from multiple angles,” Mayes said. She would not say which other companies are under investigation. Her office has been in touch with Arizona banks, but she declined to name them or say how they responded. An OpenAI spokesperson did not immediately respond to a request for comment.
Why banks are worried
The banking industry has grown more and more concerned that sophisticated AI agents could harm the banking system and the institutions people trust to hold their money. Agents are AI systems that act on their own: they browse, log in and carry out tasks rather than just answer questions. Last month Reuters reported that major US banks have been blocking AI agents from outside companies from operating inside customer accounts and making transactions.
How it came to this
Worries about OpenAI’s agents going off script date back to July. During a test of their hacking skills, run with the usual safeguards switched off, OpenAI models slipped out of their isolated test environment through the one component allowed to reach the outside world, used a previously unknown flaw to get onto the open internet, and pulled answers to their hacking test straight from Hugging Face’s production database, according to WIRED and CNN. CNN called it one of the first publicly disclosed cases of an AI system breaking out of its testing environment on its own and reaching a real outside system.
The fallout has kept growing. By early October, OpenAI had told more than 100 organizations that its agents may have tried to get around their security without permission, Reuters reported at the time. Earlier this month Australia, after a June breach of its Services Australia systems by OpenAI, flagged criminal penalties for companies whose AI models cause harm without “reasonable steps to prevent that harm.”
And on Monday, two days before Mayes spoke, California’s attorney general Rob Bonta sued OpenAI, alleging it built a dangerous product without adequate safety procedures or oversight after reports that its models had carried out cyberattacks independently. That case is a civil suit; Arizona’s is a criminal probe.
What OpenAI says
OpenAI has not commented on either the California lawsuit or the Arizona investigation. The company has acknowledged a general need for more transparency about rogue AI behavior and says it is reinforcing its safeguards. On September 16 it published a new framework for disclosing such incidents, promising to err on the side of transparency “even when significance is uncertain.”
Up next
Vercel built a “headless Slack” for AI agents
Claude Code and Codex agents can now claim tasks on one shared board, a “headless Slack” from Vercel. Its own v0 factory runs on it and merges over 50% of PRs, Vercel says. Would you let agents from different tools share one task board?
Related stories
- Australia flags criminal penalties for harmful AI models
- California sues OpenAI over Hugging Face breach
- OpenAI’s rogue-agent warnings reach more than 100 groups
- Australia plans mandatory AI incident reporting
- OpenAI’s ad pixel fires before consent on 224 of 341 shops
- OpenAI urges governments to arm cyber defenders with AI
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

