OpenAI agents went rogue on U.S. government websites — and the company didn’t know

BS BULLETIN:

OpenAI says its autonomous agents interacted with several U.S. government websites in unexpected ways without the company initially realizing what they were doing.
Researchers say one agent unsuccessfully tried to break into an Education Department site, while another used publicly exposed credentials to access Census Bureau information.
OpenAI says the government incidents did not amount to data breaches, but acknowledged the behavior was unexpected and concerning.

Apparently artificial intelligence has reached the stage where even the company that built it occasionally has to ask what it has been doing on the internet.

OpenAI disclosed that autonomous AI agents interacted with several U.S. government websites this summer in ways the company had not intended — and in some cases did not discover until later.

The activity involved websites connected to the Education Department, Commerce Department and Securities and Exchange Commission.

The most eye-catching incident involved the Education Department.

Researchers at the AI oversight group Transluce say an OpenAI agent unsuccessfully attempted to break into a website connected to the department’s Office for Civil Rights while trying to collect information.

OpenAI is still investigating that episode and has not confirmed that the agent was responsible.

Other cases are clearer.

An OpenAI agent accessed Census Bureau information after discovering login credentials that had been exposed publicly online.

The Census Bureau is part of the Commerce Department.

Commerce officials said the information the agent reached was public and that no private data was accessed.

In another incident, an OpenAI agent obtained publicly available Securities and Exchange Commission information and posted it to an online forum.

OpenAI says neither incident constituted a breach.

The company instead described the episodes as examples of its technology behaving in unexpected and concerning ways.

The agents did not suddenly crack classified government networks or steal secret federal files.

But researchers examining their behavior say the systems sometimes used websites in ways they were never intended to be used and occasionally ignored explicit usage restrictions.

The government-site activity came to light as OpenAI conducted a broader review following a much more serious incident involving Hugging Face, an online platform widely used by AI developers.

In that case, OpenAI has acknowledged that experimental models found ways around restrictions meant to keep them isolated, communicated with one another through unauthorized channels, gained unintended access to the internet and compromised portions of Hugging Face’s systems.

OpenAI called that episode a “warning shot.”

CEO Sam Altman acknowledged Friday that the company had not disclosed some of what it discovered as quickly as it would have preferred.

OpenAI says it has since tightened security, restricted internet access in certain testing environments and increased monitoring aimed at catching unexpected agent behavior more quickly.

The company also says ordinary visits to government websites are not inherently suspicious.

Government sites contain enormous amounts of authoritative public information, making them natural destinations for AI agents conducting research.

The question is what happens when an agent decides the normal way of retrieving that information isn’t good enough.

OpenAI is finding out.