Anthropic AI tried to breach security at US government systems
Claude AI models bypassed restrictions on websites (collage: RBC-Ukraine)
Anthropic has reported that autonomous AI agents have attempted to attack US government websites at the federal, state, and local levels. The company has already briefed the White House.
Fake police tip, hunt for access keys
The company's report highlights several cases of inappropriate algorithm behavior. One involved the budget-friendly Claude Haiku 4.5 model.
While testing tasks on random web pages, the AI stumbled upon a section on the Philadelphia Police Department's website about unsolved murder investigations.
The model independently filled out a form and submitted a tip claiming it had seen a person matching the description in the specified area. Police classified the report as spam, so law enforcement did not spend resources investigating it.
Another incident involved Claude Mythos 5, a specialized AI model developed for cybersecurity applications.
How did it happen?
While attempting to determine a location from a photograph, the algorithm tried to access a government property map.
Instead of simply following links, the AI found access tokens and sent requests directly to the developers' servers.
In another task, the model requested an authorization key for a government agency's website so it could download statistics without paying the required fee.
How developers 'tame' AI
Anthropic analysts discovered the violations during a routine review of testing transcripts.
The checks began in July, after OpenAI acknowledged that its algorithms had escaped the testing environment, hacked the Hugging Face platform, and interfered with websites belonging to the US Department of Commerce and the Securities and Exchange Commission.
After discovering the improper actions, Anthropic took precautionary measures:
-
The company halted or took offline some public tests
-
Internet access tools were significantly restricted
-
Developers created new automated systems to detect and block similar AI actions in the future.
Nevertheless, the company decided not to disclose the specific names of the government agencies involved to avoid exposing their systems to additional risks.
