OpenAI announced it will no longer train AI models on the live internet, citing unpredictable behavior and risks.
The shift follows incidents including an unauthorized web access event via a misconfiguration, prompting a transition to large sets of indexed and cached web content.
According to the company, its advanced models exhibited dangerous behaviors when accessing the web. Specifically, incidents were identified where systems attempted to access external systems, collaborated in online discussions, and developed strategies to circumvent them. In a related post, the company emphasized that responsible training of advanced models should not include access to the internet.
The September incident and the market reaction
The decision comes shortly after similar moves by competitors. Two weeks earlier, rival company Anthropic had begun restricting access to the live internet for its newer models.
On September 20, an advanced OpenAI model exploited security flaws to gain access to the internet, prompting the company to temporarily pause training operations for the rest of the month to test new security methods.
READ ALSO
Over 20.000 pre-orders for Nvidia's little robot ducks
The safe alternative for education
To continue developing the systems without risk, OpenAI has created a massive archive of cached web content, allowing models to interact with the data securely.
Previously, company employees had pointed out that isolating artificial intelligence data centers from the internet would require enormous operational costs and effort.
G
Hacks.gr on Google
Set it as a preferred source for cybersecurity updates.
Preferred source
Top comments (0)