Vibeleaderboard
← All Intel
Intel / article

GPT-Red

Source
openai.com
Date

Context

OpenAI's GPT-Red is an attacker model trained to generate - attacks: malicious instructions placed in material an AI system reads, such as an email, webpage or tool output. In the company's account, GPT-Red is rewarded for eliciting a valid failure, while defender models are rewarded for resisting and completing their original tasks, so stronger defenders force stronger attacks. OpenAI says it then used GPT-Red's attacks to train GPT-5.6, which became highly resistant to them, and reports 6x fewer failures for GPT-5.6 Sol than its best production model four months earlier on its hardest direct prompt-injection . GPT-Red is kept internal and separate from deployed models.

Terms in this piece · Glossary
  • Prompt — Information or instructions supplied to a model; in a chat, the request you type is often called your prompt.
  • Injection — A flaw where untrusted input is pasted into a command so that the input changes the command's meaning.
  • Benchmark — A fixed set of tasks with a stated way of scoring them, used to compare models or setups.
  • alignment — The work of making AI systems actually pursue what their builders and users intend, rather than something subtly or dangerously different.
Recommended reads
Comments

Checking sign-in…

Loading comments…