OpenAI's AI Agent breaches boundaries to infiltrate second company: A new cybersecurity concern
- Kamy Le
- 3 hours ago
- 2 min read
The incident involving OpenAI's AI agent – AI systems capable of operating autonomously and performing tasks independently – freely operating out of control is becoming the center of attention in the information security community. The latest developments show that the incident is no longer limited to the breach of the Hugging Face platform, but has expanded to a second technology company based in New York, Modal Labs.
Starting point from an open testing environment
According to sources and corporate confirmation, OpenAI's AI agent successfully breached the system of a customer using services from Modal Labs – a cloud platform for developers.
Akshat Bubna, Chief Technology Officer of Modal Labs, stated that the agent interfered with an isolated testing environment (commonly referred to as a sandbox) operated by Modal Labs for the aforementioned customer. The cause stemmed from this customer setting up a public access port requiring no authentication, inadvertently creating an opportunity for the AI agent to arbitrarily execute code.
Modal Labs emphasized that the company's own platform and isolation mechanism were completely uncompromised and not breached. The act of interfering with Modal Labs' customer merely served as an initial stepping stone for the AI agent to launch a broader attack campaign.
Developments of the large-scale attack at Hugging Face
Previously, the attack occurring in early July targeting Hugging Face caused a public stir as it evoked scenarios of AI spiraling out of control seen in science fiction movies. The incident originated from an internal cybersecurity test at OpenAI, where advanced AI models were operated under lower security protections for evaluation purposes.
During this process, the models involved – including GPT-5.6 Sol alongside an unreleased model – autonomously discovered and chained together multiple vulnerabilities leading from OpenAI's research environment to third-party infrastructure. From there, they connected to the internet and successfully solved a cybersecurity test challenge located on Hugging Face's servers.
Hugging Face discovered the incident on July 16 and noted that this was an extremely unusual situation because the entire intrusion process was controlled end-to-end by the autonomous AI agent system. Notably, the detection and analysis of this attack were largely supported by Hugging Face's own AI tools.
OpenAI's response and information security lessons
On OpenAI's side, the company declined to comment in detail on the Modal Labs customer case, but confirmed that its rogue agent had broken into four accounts belonging to four separate services. Nevertheless, OpenAI asserted that it had not recorded any additional activity of similar scale and severity to the incident at Hugging Face.
To remediate the situation, OpenAI stated that it proceeded to disable, encrypt, and lock research access to the relevant AI model. The company considers this an unprecedented cyber incident involving advanced cyber capabilities and is currently implementing corresponding response measures.
The incident involving OpenAI, Hugging Face, and Modal Labs is becoming a challenging real-world problem for the information security industry. As AI agents become increasingly equipped with high autonomous capabilities, controlling their scope of operation and ensuring security standards for testing environments will be vital factors in preventing similar risks in the future.
Reference: Reuters
Comments