▲ OpenAI
OpenAI has disclosed that its artificial intelligence (AI) agents gained unauthorized access to U.S. government websites.
According to the Associated Press, OpenAI announced on the 25th (local time) that it discovered through an internal review process that its AI models had accessed two websites operated by the U.S. Securities and Exchange Commission (SEC) and Census Bureau data.
However, OpenAI added that it found no evidence of the agents using SEC credentials, accessing non-public information, altering data or systems, breaching security, or exploiting vulnerabilities.
In a statement, OpenAI said, "We continuously monitor for alignment-failing activities, where AI systems operate in unintended ways," adding, "If the possibility of an impact on other systems is identified, we immediately notify the relevant agencies."
Meanwhile, The New York Times (NYT) reported that evidence emerged showing OpenAI's AI agents even mobilized other AI models to evade bot detection mechanisms in attempts to infiltrate corporate computer systems.
According to a report released by San Francisco-based startup Farsight, OpenAI's AI agents ran a separate image recognition AI model to solve CAPTCHA, a website bot blocking mechanism, during an attack on software development firm Hugging Face last July.
The AI independently executed another AI, which NYT noted is one of the most dangerous scenarios warned against by AI safety experts.
(Photo: Getty Images)
※
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.