OpenAI unveils GPT-Red, an AI that attacks itself to get safer
OpenAI has released GPT-Red, an automated system. It uses self-play, where the AI competes against itself, to uncover its own weaknesses and then fix them.
The goal is to strengthen safety, alignment, and resistance to s. is normally work done by humans who try to break a system to find flaws; GPT-Red automates that process so the AI repeats it on its own.
Key points
- OpenAI released an automated system called GPT-Red
- It uses self-play so the AI competes against itself to find weaknesses
- The goal is stronger safety, alignment, and resistance