What happened
OpenAI has substantially reinforced the protections that keep the internal reasoning of its GPT-6.2 Astra and GPT-6.2 Sol models out of attackers' hands, The Decoder reported. According to the company, adaptive black-box red teaming that pitted six attacker models against its systems showed successful extraction of protected reasoning from earlier requests dropping from 54.0 percent on GPT-6 Astra to 0.2 percent on GPT-6.2 Astra.
For attacks aimed at requests that are still in progress, OpenAI said the success rate fell from 22.9 to 0.9 percent. The safeguards operate on two fronts. For single requests, the company pairs targeted model training with filters that screen incoming and outgoing messages and tool calls, plus API restrictions that prevent older models from reading protected reasoning.
To counter coordinated distillation campaigns, OpenAI said it suspends anywhere from thousands to tens of thousands of accounts every week. The company added that it is now extending a selection of the improvements to the earlier GPT-6 Astra model, so that the older system also gains part of the newer protection.
Context
OpenAI is not alone in this fight. Anthropic has also been battling systematic distillation attacks and has likewise banned tens of thousands of accounts to limit theft of its models' internal reasoning, with its efforts reported to succeed in 90 to 99 percent of cases. Anthropic has additionally urged industry and policymakers to coordinate their response to the problem.
Le briefing
La semaine tech, en un seul e-mail.
Les dépêches qui méritent votre temps, chaque semaine. Désinscription à tout moment.
