OpenAI has announced significant improvements in the security of its AI models, specifically GPT-6.2 Astra and GPT-6.2 Sol. In recent adaptive red-teaming tests, the success rate for attackers attempting to extract internal reasoning from these models plummeted from 54 percent to just 0.2 percent. This marks a substantial enhancement in the protection of the models’ reasoning processes.
To further bolster security, OpenAI is actively suspending thousands of accounts weekly that are involved in coordinated distillation campaigns aimed at exploiting the models. These measures underscore OpenAI’s commitment to maintaining the integrity and security of its AI technologies against malicious attempts to reverse-engineer their reasoning capabilities.
Compiled automatically by the Tech AI Newsdesk from public AI-news sources and summarised in our own words.