OpenAI Uses AI Self-Play to Attack Its Own Models
The Decoder · rss · 2026-07-16
OpenAI's internal GPT-Red model, trained via self-play, successfully launched attacks in test scenarios 84% of the time, compared to a mere 13% success rate for human red team members.
These test results are being directly applied to enhance the safety and defense capabilities of models like GPT-5.6 Sol.
More from Models
- GPT-6 Astra beats Factorio with enemies in 44 in-game hours at ~$4,500 API cost — liminal_bardo · 2026-09-11
- 105 hidden bugs, 2 repos: DeepSeek V4.1 Flash fixes 24 at $1.80 vs Opus 5's 27 at $51.33 — ChartsJournalX · 2026-09-11
- awesome-llm-leaderboards: an open-source directory of LLM leaderboards, pricing tables, comparison tools — Last_Establishment_1 · 2026-09-11
- Anthropic claims it works to keep eval environments unidentifiable to models — MaxKannen · 2026-09-11
- Nex N2.5 Pro released on Hugging Face with 407GB of weights — jinnyjuice · 2026-09-11
- RoMa v2 image matching model unveiled in the usual black poster — ducha_aiki · 2026-09-11