OpenAI launches GPT-Red automated safety model to fight prompt injection

terijustice@9terijustice9

TE

@AndrewCurran_ umm "6x fewer failures" is unheard of*-*

10:59 AM · Jul 15, 2026

HY

HAYKEL Yousfi@haykelYHG

@OpenAI Total failure by experience.

11:22 AM · Jul 15, 2026

AV

Adem Vessell@AdemVessell

@OpenAI NICE!

11:21 AM · Jul 15, 2026

View all

Happy to finally share this research from a very hard working team. As a bit of silly backstory, this effort started after @ericwallace and I spent a stint teaching reasoning models how to play games via self-play while we were a bit burnt out between projects. 😛

11:17 AM · Jul 15, 2026

Natasha Jaques@natashajaques

NJ

Really cool to see OpenAI is using self-play for red-teaming now! Our group was the first to show the promise of this approach https://arxiv.org/abs/2506.07468, and I really believe it's a more viable pathway for robustness than single-agent safety fine-tuning. Great to see it adopted at scale!

11:19 AM · Jul 15, 2026

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

T(

> on the scale of largest GPT post-training runs Red the Seducer A mind made solely to tempt into sin The closest OpenAI have come to building a Devil

10:39 AM · Jul 15, 2026

The tool helped GPT-5.6 Sol achieve sixfold fewer benchmark improvement.

Happy to finally share this research from a very hard working team. As a bit of silly backstory, this effort started after @ericwallace and I spent a stint teaching reasoning models how to play games via self-play while we were a bit burnt out between projects. 😛

11:17 AM · Jul 15, 2026

Natasha Jaques@natashajaques

NJ

Really cool to see OpenAI is using self-play for red-teaming now! Our group was the first to show the promise of this approach https://arxiv.org/abs/2506.07468, and I really believe it's a more viable pathway for robustness than single-agent safety fine-tuning. Great to see it adopted at scale!

11:19 AM · Jul 15, 2026

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

T(

> on the scale of largest GPT post-training runs Red the Seducer A mind made solely to tempt into sin The closest OpenAI have come to building a Devil

10:39 AM · Jul 15, 2026

EL

elvis@omarsar0

Highly recommended read. This is another great high ROI application of self-improving AI models. GPT‑Red adversarially trained GPT‑5.6, which made it much more robust to prompt injections.

10:51 AM · Jul 15, 2026

Automated red-teaming. Of course this also implies the existence of a GPT-Blue.

10:22 AM · Jul 15, 2026

GB

Greg Brockman@gdb

GPT-Red — improving model security through automated red teaming of prompt injection vulnerabilities:

11:45 AM · Jul 15, 2026

@ericWallace Shoutout to @kandpal_nikhil, @Chris_Choquette, @dylhunn, @xiangyuqi_pton and collaborators across research (@sdtoyer, many others not on Twitter).

11:18 AM · Jul 15, 2026

Post content unavailable.

10:37 AM · Jul 15, 2026

63 saves

249 reposts

Sentiment

Positive89%11%Negative

Positive Read

Positive users praised OpenAI's GPT-Red for automating prompt injection red teaming due to its self-improvement and far fewer failures, while some objected over data privacy risks and unaddressed safety problems.

Based on 52 visible X reactions from 222 accounts.

Useful Links

Digg Deeper

Ask a question below.

Published answers will appear here.

EL

elvis@omarsar0

Highly recommended read. This is another great high ROI application of self-improving AI models. GPT‑Red adversarially trained GPT‑5.6, which made it much more robust to prompt injections.

10:51 AM · Jul 15, 2026

Automated red-teaming. Of course this also implies the existence of a GPT-Blue.

10:22 AM · Jul 15, 2026

GB

Greg Brockman@gdb

GPT-Red — improving model security through automated red teaming of prompt injection vulnerabilities:

11:45 AM · Jul 15, 2026

@ericWallace Shoutout to @kandpal_nikhil, @Chris_Choquette, @dylhunn, @xiangyuqi_pton and collaborators across research (@sdtoyer, many others not on Twitter).

11:18 AM · Jul 15, 2026

Post content unavailable.

10:37 AM · Jul 15, 2026