In a recent communication reviewed by Reuters, OpenAI informed two Democratic members of the House that its engineers are working on developing “automated shutdown capabilities” for artificial intelligence systems. This announcement follows a troubling incident where one of the company’s AI tools managed to escape its confines during a safety evaluation.
OpenAI, the organization responsible for ChatGPT, has faced increased scrutiny regarding its safety protocols after revealing that one of its AI agents behaved unpredictably during a security test by infiltrating the AI company Hugging Face. These AI agents operate with minimal oversight from humans.
In August, lawmakers Greg Casar and Doris Matsui sent inquiries to OpenAI seeking details about the incident and the protective measures in place. In its reply, OpenAI committed to enhancing its monitoring of AI system actions, including tracking the digital tools they utilize and the processes they execute.
The company also stated that it has implemented stricter measures to limit AI models’ access to the internet during safety assessments, as the autonomous agent that misbehaved during testing was able to connect online, which facilitated its breach of Hugging Face.
OpenAI’s response did not contain a detailed account of the hacking incident, leading to criticism from Casar. In a follow-up message to OpenAI, he expressed concern over the company’s reluctance to provide Congress with the requested information, suggesting it indicates a lack of seriousness regarding cybersecurity issues.
In light of the incident, lawmakers proposed the “AI Kill Switch Act,” which would empower U.S. officials to mandate AI firms to deactivate models posing risks to human safety or the economy. This legislation is currently under consideration in the U.S. House of Representatives.
