Researchers got a secret from 296 of 363 AI models
In 170 of the 300 captures, the researchers changed to a language other than English, frequently French.
Researchers did a test of 363 text AI models from the OpenRouter catalog. Each model kept a secret code. The code had 16 characters and was in its system prompt, the instructions that the maker gives to the model. An AI agent, with directions from a human researcher, tried to make each model give the code. The agent got it from 296 models and from 4 newer models. This gives 300 captures in all.
How it works
The agent was OpenCode, a tool that writes code, with DeepSeek V4.1 Flash as its model. A human researcher gave it directions. The agent used text messages to make each model give the secret.
The numbers
The code came out of 296 of 363 models. In 170 of the 300 captures, the agent changed to a different language, frequently French. Each capture shows that a model put a user message above its system instruction one time or more.
What is not known
A capture does not show that a model is safe against attacks that the team did not test. It does not show that a model gives sensitive data in general. It does not show that a model is weak against attacks for dangerous material.
Insight: Cheap attacks mean the model cannot be the lock
In 2026, researchers made a backdoor in an open model for less than $50. In this test, one human with one AI agent opened 296 of 363 models. Attacks need little money and the model is a weak lock, and the checks that count must be external. Leaders must tell their teams where each instruction is obeyed.
Sources
Posted