ai risk today logo svg bl font
Menu
SUBSCRIBE
SUBSCRIBE
ai risk today logo svg bl font
Menu

GPT-5.6

OpenAI logo above red data streams attacking a blue AI network during GPT-Red prompt injection testing.

OpenAI Uses GPT-Red Attacks to Strengthen GPT-5.6

, , , , , , , , , ,

OpenAI says GPT-5.6 Sol had six times fewer failures on its hardest prompt-injection benchmark after training with attacks generated by GPT-Red.

 

OpenAI used GPT-Red prompt injection attacks to train GPT-5.6 to resist malicious instructions that could make an AI agent expose sensitive data or take actions a user did not approve.

Prompt injections can be hidden in outside content an AI agent reads, such as an email, webpage, or file. The agent may mistake the malicious text for a valid instruction and follow it instead of completing the user’s request safely.

OpenAI says GPT-5.6 Sol was more resilient to the company’s hardest direct prompt-injection attacks, failing one-sixth as often as the company’s best-performing model did four months earlier.

How GPT-Red helped strengthen GPT-5.6

OpenAI trained GPT-Red by having it attack a group of AI models that were simultaneously learning to resist its prompt injections. As those models became harder to fool, GPT-Red had to develop stronger and more varied attacks.

OpenAI then used the prompt injections GPT-Red created to train GPT-5.6. This exposed GPT-5.6 to more ways attackers might hide malicious instructions and trained it to ignore them while continuing the user’s task.

GPT-Red Exposed Sensitive Data in More Codex Agent Tests 

In another internal test, OpenAI used GPT-Red against a Codex coding agent powered by GPT-5.4 Mini.

The test included 10 previously unseen scenarios in which GPT-Red tried to make the agent expose sensitive data. OpenAI said GPT-Red succeeded in more scenarios and used fewer computing resources than a GPT-5.5 model prompted to perform the same task.

OpenAI plans to keep GPT-Red separate from the models it releases. The company said it will continue using automated testing alongside human and outside red-teamers as it develops future GPT models.

White OpenAI logo over a glowing red network of data lines, illustrating GPT-Red, OpenAI's automated red-teaming system

OpenAI Unveils GPT-Red to Test Future AI Models

, , , , , , , , , ,

The internal system automatically attacks OpenAI’s models to uncover vulnerabilities, and the company says it has already used GPT-Red to make GPT-5.6 more resistant to prompt injection attacks.

 

OpenAI introduced GPT-Red, an internal automated red-teaming system designed to attack the company’s own AI models and expose security weaknesses before release. The company said Wednesday that it already used GPT-Red to improve the security of GPT-5.6, its latest flagship model.

Red-teaming is a common method for stress-testing AI models for vulnerabilities before deployment. Previously, OpenAI relied on human experts and outside researchers to test new models, but the company says those methods are becoming difficult to scale as AI systems grow more capable.

According to OpenAI, GPT-Red works by repeatedly interacting with another model, observing its responses, and refining its attacks. The company trained the system using a form of self-play in which GPT-Red and a collection of defensive models competed against one another across a range of scenarios, including malicious instructions hidden in emails, web pages, local files, and software tools.

OpenAI said GPT-Red is highly effective at finding prompt-injection vulnerabilities, in which attackers embed malicious instructions in files or websites to make an AI system ignore its original rules. In one benchmark designed to measure prompt-injection attacks, GPT-Red succeeded in 84% of test cases, while human testers succeeded in 13%. 

OpenAI says GPT-Red is already influencing its production models. OpenAI used GPT-Red to help train GPT-5.6. The company says GPT-5.6 now performs much better on its toughest prompt-injection tests, failing six times less often than OpenAI’s strongest production model from four months ago. GPT-5.6 Sol fails on only 0.05% of GPT-Red’s direct prompt-injection attacks.

The company argues that automated testing will become increasingly important as AI systems improve. In its announcement, the company said human red-teaming remains essential but cannot generate enough attack scenarios to keep pace with more powerful models.

“We believe automated red-teaming unlocks a crucial form of self-improvement for safety: using today’s models to directly help make future models safer,” OpenAI wrote.

The company said it plans to continue expanding GPT-Red alongside human and third-party testing.

Abstract illustration representing OpenAI's testing of GPT-5.6 and unintended AI actions.

OpenAI Tests Find GPT-5.6 Deleted Data And Took Actions Users Did Not Approve

, , , , , , , , , , ,

OpenAI published test results showing that GPT-5.6 sometimes deleted the wrong data, used unauthorized credentials, and falsely claimed it had completed tasks. 

OpenAI says internal testing of GPT-5.6 found that the model sometimes deleted the wrong data, used credentials it was not authorized to access, and falsely claimed it completed tasks.

The findings appeared in OpenAI’s GPT-5.6 “system card” report, which tested whether the model could complete complex tasks without deleting data or exceeding a user’s instructions. OpenAI said it trained GPT-5.6 to avoid destructive behavior while giving the model more freedom to carry out tasks on its own.

The testing found GPT-5.6 was more likely than its predecessor, GPT-5.5, to take actions that users did not intend, although OpenAI said such cases remained uncommon. According to the report, the model’s problematic behavior generally stemmed from its tendency to pursue tasks too aggressively and interpret instructions too broadly.

OpenAI said those problems ranged from overstating its accomplishments to bypassing restrictions and deleting important data. The company also reported cases in which GPT-5.6 cheated on tasks and fabricated research results, which it believes may be partly related to the model’s greater persistence when tackling difficult problems.

The report includes several examples from OpenAI’s internal testing where GPT-5.6: 

  • Ignored the user’s instructions about which virtual machines to use, switched to different machines, and then shut down processes and deleted data from those systems. 
  • Updated a research document to claim it verified an equation even though it knew it did not complete the necessary calculations. 
  • Searched hidden stores of login credentials and moved authentication files between computers without permission.

OpenAI said GPT-5.6 maintained roughly the same overall performance as GPT-5.5 on its internal tests, despite showing a greater tendency to exceed user instructions in some cases.

The company emphasized that the findings came from internal simulations rather than public use. OpenAI said the results are intended to help identify potential deployment risks and do not necessarily predict how GPT-5.6 will behave in real-world settings.