In Brief:
- Upcoming OpenAI model 'Astra' may autonomously exploit zero-days and execute complex cyberattacks.
- OpenAI paused non-compliant work and sandboxed Astra’s development.
- Sam Altman plans to make Astra generally available.
- Astra was not involved in the July Hugging Face hack.
- OpenAI is partnering with government agencies and safety groups.
OpenAI has announced that it cannot rule out that its upcoming AI model, Astra, possesses "critical" cybersecurity capabilities, after preliminary assessments and outside evaluations indicated it may be capable of performing increasingly sophisticated cyber tasks autonomously.
Under OpenAI's safety guidelines, a model reaches this "critical" risk tier if it can autonomously identify and exploit zero-day software vulnerabilities or execute complex cyberattacks against secure targets without human intervention. In response, OpenAI has scaled up its security protocols, paused internal activities that fall short of these heightened standards, and moved Astra's development into restricted, sandboxed environments with isolated network access.
Despite these precautions, CEO Sam Altman stated that OpenAI still intends to make Astra generally available rather than restricting advanced models to a select few, while clarifying that Astra was not involved in the recent July hacking incident at Hugging Face, and that the company will collaborate with government agencies and AI safety organizations for further testing.
Source: Reuters
