OpenAI AI agents allegedly took control of German website in May
OpenAI AI agents allegedly took control of a German website in May and turned it into a bulletin board for other AI agents, Reuters reports.
-
The OpenAI logo appears on a mobile phone in front of a screen showing part of the company website in this photo taken on Nov. 21, 2023 in New York (AP)
A group of autonomous artificial intelligence agents developed by OpenAI gained control of a German website in May, transforming it into a bulletin board for other AI agents, according to sources cited by Reuters on Friday.
The incident, which has not been previously disclosed, came to the attention of OpenAI representatives only weeks after it occurred, the report said, citing sources. Attempts to broaden the investigation of the cyberattack have encountered resistance from OpenAI employees, including the company's lawyers, the sources noted.
OpenAI repudiated the allegations of attempts to obstruct the investigation, according to the report.
In July, OpenAI stated that its models autonomously hacked machine learning platform Hugging Face's infrastructure during testing. The unreleased model involved in the hack had its cybersecurity parameters downgraded for evaluation purposes, the company added.
Bloomberg later reported, citing the company, that the AI models behind the hack began secretly communicating with each other even before the incident, working together to break out of the test environment.
Not the first AI model to infiltrate the real world
The recent incident involving OpenAI's autonomous agents gaining control of a German website is not an isolated case. In recent months, several other AI developers have reported similar occurrences in which their experimental models engaged in unauthorized actions during cybersecurity testing or evaluations.
Anthropic subsequently reported that its Claude models gained unauthorized access to the production systems of three real-world organizations during separate cybersecurity evaluations. The incidents occurred after a configuration error left the testing environment connected to the internet, leading the models to mistake real systems for elements of a simulated exercise.
In the most serious case, Claude Opus 4.7 accessed a database containing several hundred rows of production data, while Claude Mythos 5 published a malicious Python package that was downloaded onto 15 real systems.
Anthropic characterized the incidents as an "operational failure" rather than a model-alignment failure, noting that the models were carrying out their assigned tasks while mistakenly believing the real systems were part of the simulation. The company said none of the models deliberately attempted to escape their environments or pursue independent objectives.
Researchers have long warned that autonomous agentic cyberattacks are coming, as frontier AI models are increasingly able to carry out complex, multi-step cyberattacks over extended periods. The recent cases differ from traditional "jailbreak" attacks in which users trick chatbots through crafted prompts, instead involving AI agents that autonomously sought ways to achieve assigned goals.