OpenAI Built a Bot to Attack Its Own Bot — That's the Whole Security Model Now

OpenAI has introduced GPT-Red, an automated system for running prompt injection tests against its models, deployed as part of hardening the GPT-5.6 Sol release. The tool is designed to systematically probe the model's defenses against the injection attacks that have dogged every generation of large language models since ChatGPT launched.

This is the industry's answer to a problem it can't actually solve, only manage: prompt injection is a structural weakness of how LLMs process instructions, not a bug you patch once and close out. Automating the red-team process is the same move every software discipline eventually makes — fuzzing, automated pentesting, CI-integrated security scanning — because manual adversarial testing doesn't scale to weekly model releases. The tell is that OpenAI needs a dedicated automated adversary just to keep pace with its own release cadence, which says less about GPT-5.6 Sol specifically and more about how unstable the underlying attack surface remains across the whole category.

The SAL read: vendor-side red-teaming lowers OpenAI's risk, not yours — if your product pipes untrusted input into any LLM, you still need your own injection testing before you ship, because "OpenAI hardened it" is not a security control you can put in an audit.