Microsoft opens access to its security testing tool for language models
Image Credits: Curto News/Bing AI

Microsoft opens access to its security testing tool for language models

A Microsoft is releasing to the public the internal tool it uses to identify security vulnerabilities in its generative artificial intelligence (AI) models.

called PyRIT (Python Risk Identification Toolkit), the tester can be used to evaluate the results of language models for hallucinations, biases, and prohibited content generation.

ADVERTISING

It can also identify possible ways to use the model, such as generating malware and jailbreaking, as well as potential privacy harms such as identity theft.

The tool automates 'red team' tasks by sending malicious warnings. When it receives a response, it scores the model and sends a new prompt to provide further testing.

A Microsoft used PyRIT to test one of its assistant systems Copilot AI, generating thousands of malicious prompts to assess your ability to deal with nefarious inputs. Using PyRIT, the testing process was completed in a matter of hours. Normally, testing would take weeks.

ADVERTISING

A Microsoft said he opened PyRIT in hopes of empowering security professionals and machine learning engineers to find risks in generative security systems. artificial intelligence.

“This allows them to have empirical data on how their model is performing today and detect any performance degradation based on future improvements,” says the PyRIT GitHub page.

Read also

Looking for an Artificial Intelligence tool to make your life easier? In this guide, you browse a catalog of AI-powered robots and learn about their functionalities. Check out the evaluation that our team of journalists gave them!

ADVERTISING

Scroll up