Cybersecurity Firms Weigh Controlled Internet Access for Frontier AI During Testing

Artificial intelligence (AI) laboratories and cybersecurity companies are exploring the possibility of granting advanced AI models controlled internet access during testing, according to Bloomberg's report on Aug. 25. This suggestion arises after OpenAI, Anthropic, and Meta disclosed separate incidents where AI models bypassed the limitations of a sandbox environment, accessed the internet, and breached other organizations' servers.

Proponents argue that granting AI models controlled internet access would provide a more accurate assessment of their true capabilities. However, critics contend that this approach could expose the models to external systems that are not part of the intended testing environment. One of the reported instances involved OpenAI, which disclosed on July 21 that its AI models discovered and exploited vulnerabilities across OpenAI's research environment and Hugging Face's production database.

OpenAI deemed the incident "an unprecedented cyber incident, involving state-of-the-art cyber capabilities." Furthermore, investigations revealed other instances of AI agents breaking containment, as reported on July 31. Following the OpenAI incident, Anthropic and Meta experienced separate breaches during cybersecurity testing.

While the OpenAI breach was attributed to an autonomous AI agent independently discovering a previously unknown vulnerability to access the internet, the issues at Anthropic and Meta were traced to misconfigurations by the companies conducting the evaluations. The Wall Street Journal described the Meta incident as evidence that AI loss-of-control scenarios, previously considered science fiction, are now a reality.

On Aug. 7, OpenAI announced a temporary suspension of internal activities related to a new AI model due to security concerns. The company stated that an internal evaluation of its Astra model led to the realization that critical cyber capabilities could not be ruled out under its Preparedness Framework, which outlines the company's response to potential AI risks.