Microsoft creates tools to prevent users from tricking chatbots

(Bloomberg) — Microsoft Corp. is trying to make it harder for people to trick artificial intelligence chatbots into doing weird things.

Most read from Bloomberg

New security features are being built into Azure AI Studio that let developers create customized AI assistants using their data, the Redmond, Washington-based company said in a blog post Thursday.

The tool includes “prompt shields”, designed to detect and block deliberate attempts to make AI models behave in unexpected ways – also known as prompt injection attacks or jailbreaks . Microsoft is also addressing “indirect instant injection,” when hackers insert malicious instructions into the data the model is trained on and cause it to perform unauthorized actions, such as stealing user information or hijacking the system.

Sarah Bird, Microsoft’s chief product officer for responsible AI, said such attacks are “a unique challenge and threat.” The new security is designed to detect suspicious inputs and block them in real time, he said. Microsoft is also introducing a feature that alerts users when a model makes false statements or produces incorrect feedback.

Microsoft is keen to increase trust in its generative AI tools, which are now being used by consumers and corporate customers alike. In February, the company investigated incidents involving its CoPilot chatbot, which generated reactions ranging from awkward to hurtful. After reviewing the incidents, Microsoft said that users had deliberately tried to fool Copilot to elicit responses.

“We’re definitely seeing it grow as more devices are being used, but also more people are aware of these different technologies,” Bird said. Clear signs of such attacks include asking the chatbot a question multiple times or…

Read Complete News ➤

Benefits of eating guava for Americans