OpenAI reveals GPT-Red to secure GPT-5.6 and GPT-5.6 Sol
OpenAI just revealed GPT-Red, a new tool that helps spot and fix weaknesses in its AI models before they go public.
Basically, GPT-Red acts like a security check, running simulated attacks on the models to catch flaws early.
It played a major role in improving the security of OpenAI's latest releases, GPT-5.6 and GPT-5.6 Sol, which were released earlier this week.
GPT-Red crafts prompts to bypass safety
Think of GPT-Red as an automated "red team" that tries to break the AI by crafting tricky prompts and finding ways around its safety rules, kind of like ethical hacking but for AI.
It trains itself using self-play reinforcement learning: it gets rewarded when it finds gaps in defenses while defender AIs learn to block these attacks.
OpenAI says this has made its models much tougher over the past six months.
More technical details are coming soon if you're curious about the nerdy stuff.