OpenAI has issued a new warning about its upcoming AI model Astra following a series of concerning incidents involving cybersecurity breaches, including an incident linked to Hugging Face in July.
According to the tech company, it cannot rule out the possibility that Astra may have “critical cybersecurity capabilities.” As a result, OpenAI paused some internal development work and activated its safety protocols.
Under OpenAI’s safety guidelines, a model reaches the critical level if it can independently identify and exploit real-world software vulnerabilities. In a worst-case scenario, such an AI system could potentially carry out cyberattacks against highly secure targets without human assistance.
The report comes as autonomous AI agents have attracted global attention because of their ability to bypass safety testing measures. OpenAI is not the only company facing such concerns. Other companies, including Anthropic, Meta and Chinese AI startup Moonshot, have also faced questions about their cybersecurity standards.
According to a report from OpenAI, the maker of ChatGPT, Astra’s advanced capabilities could allow the model to carry out high-level cyberattacks autonomously.
“While we continue to benchmark and assess this model, our preliminary evaluations indicate strong enough performance that we cannot rule out ‘critical’ capability level at this time,” OpenAI said.
After identifying these concerns, OpenAI strengthened its security controls and paused some internal activities. For Astra’s further development, the company plans to use an isolated testing environment with “restricted network access and sandboxed execution.”
In a post on X, Sam Altman said OpenAI was working to make Astra generally available because the company does “not think it is a good strategy to keep powerful models to a chosen few.” He added that, because of the model’s cyber capabilities, the company needs “a little bit longer” to make it safe, while expressing hope that the wait would not be too long.
