OpenAI said on July 15 that it built GPT-Red, an internal automated red-teaming model that tests AI systems for prompt injection and helps adversarially train GPT-5.6 before wider deployment.

The company said GPT-5.6 Sol had 6x fewer failures on its hardest direct prompt-injection benchmark than its best production model four months earlier.

OpenAI also said GPT-Red outperformed human red-teamers in a separate test.

In a simulated agent that closely mirrored a production AI vending machine from Andon Labs, a US-based AI evaluation startup, GPT-Red changed prices and canceled another customer’s order before OpenAI disclosed the flaws.

In a system card, OpenAI described GPT-5.6 as a three-model reasoning family.

The company said its prompt-injection evaluation tests adversarial instructions embedded in connector outputs, including search and function-calling attacks.

Comparisons elsewhere in the card were mainly against recent GPT-5.5 snapshots.

Prompt injection happens when malicious instructions in prompts or external content change a model’s behavior in unintended ways.

OWASP, short for the Open Worldwide Application Security Project, ranked it as the top large language model application risk and recommended adversarial testing.

Andon’s Vending-Bench 2 measured long-running agent performance in a simulated vending business rather than security.

The company separately said its office vending machines are internal deployments.

🔗 Source: OpenAI

Recent OpenAI developments