BREAKING
OpenAI Unveils GPT-Red
Self-Play Training Loop
1
Red-teamer attacks
↓
2
Injection succeeds
↓
3
Defender hardened
0
x
failure rate cut
0
%
new failure rate
GPT-Red vs Human Testers
GPT-Red
84
Humans
13
Internal-Only, Not Released
Pre-Print Due This Week
AI NEWS BLITZ
OpenAI just revealed GPT-Red, an automated red teamer built to hunt prompt-injection flaws.