ModelsSafetyBig tech

OpenAI makes it much harder for attackers to extract how its latest AI models think

via The Decoder · 11 views · 1d ago

AI models just got a whole lot tougher to crack. OpenAI has dramatically boosted safeguards on its latest models, GPT-6.2 Astra and Sol. Previously, attackers could extract internal reasoning from these AIs more than half the time, succeeding in 54% of tests. Now, that success rate has plummeted to a mere 0.2%. This isn't just a technical tweak; it's a massive leap in AI security. Making it nearly impossible to probe an AI's internal thought processes is vital for preventing misuse and ensuring these powerful systems remain safe. It means better protection against those seeking to exploit advanced AI.

Read the original story →

← Back to the feed

Loading comments…

Log in to join the conversation.