OpenAI just rolled out big security upgrades for its GPT-6.2 Astra model, aiming to stop people from stealing how the AI thinks, something that could be risky in cybersecurity and beyond.
Thanks to these changes, the success rate of these attacks has dropped from 54% all the way down to just 0.2%.
OpenAI cuts reasoning extraction to 0.9%
OpenAI also made it much harder to pull out reasoning from live requests, cutting success rates from 22.9% to 0.9% according to internal adaptive evaluations involving automated red-teaming.
OpenAI detected multiple unauthorized distillation campaigns on its platform this year, with weekly enforcement actions ranging from thousands to tens of thousands of accounts, and is now sharing warning signs with other AI companies and urging developers, partners, and governments to work together to keep advanced AI tech safe from misuse.