OpenAI says it uses an automated red-teaming system, GPT-Red, to strengthen GPT-5.6 against prompt injection attacks. According to the outlets, GPT-Red is designed to mimic the workflow of a human red teamer: it supplies prompts to a target model, observes the model’s responses, and iterates to probe for weaknesses. In the described setup, the system is trained using self-play reinforcement learning, where GPT-Red and defender models learn together across many different scenarios. Across this process, GPT-Red receives reward when it can elicit a “valid failure,” such as behavior consistent with a successful prompt-injection attempt, while defender models are trained to mitigate those outcomes. One outlet also reports that GPT-Red performs competitively against human red teamers on prompt injection testing, suggesting the automated approach can find vulnerabilities effectively. OpenAI says these findings from GPT-Red are used to make GPT-5.6 more resistant to prompt injection.
OpenAI says GPT-Red red-teaming improves GPT-5.6 resistance to prompt injection
OpenAI says it uses an automated red-teaming system, GPT-Red, to strengthen GPT-5.6 against prompt injection attacks. According to the outlets, GPT-Red is designed to mimic the workflow of a human red...
- OpenAI introduces an automated red-teaming model called GPT-Red.
- GPT-Red probes prompt injection weaknesses by sending prompts and monitoring responses.
- The system is trained using self-play reinforcement learning alongside defender models.
- OpenAI says results from GPT-Red are used to improve GPT-5.6’s resistance to prompt injection attacks.
- One report says GPT-Red performs better than human red teamers on a prompt injection test.
GPT-Red is an automated red-teaming model that OpenAI trains to find prompt injection weaknesses. It works the way a human red-teamer does. It sends a prompt, watches how a GPT model responds, and iterates toward a goal such as a successful data exfiltration. Training runs on self-play reinforcement learning, with GPT-Red and a set of defender models learning at the same time across many scenarios. The attacker earns reward for eliciting a valid failure. The … More → The post GPT-Red beat human red teamers on a prompt injection test appeared first on Help Net Security.
3 hours agoOpenAI said its new automated red-teaming model, GPT-Red, uncovered vulnerabilities that were used to make GPT-5.6 more resistant to prompt injection attacks.
10 hours ago
Microsoft coaches sellers on competing with Anthropic and OpenAI
Microsoft is instructing its sales teams on how to compete more aggressively with artificial intelligence companies Anth...
Paresh Rawal regrets calling Anupama Chopra “Miss Irrelevant” after Dhurandhar review
Veteran actor Paresh Rawal says he regrets a social media remark aimed at film critic Anupama Chopra following her revie...
JPMorgan CEO warns that Anthropic’s Mythos AI could pose risks if broadly released
JPMorgan Chase CEO Jamie Dimon says he is concerned about Anthropic’s most powerful AI model, “Mythos,” especially if it...