Microsoft promises better Copilot experience against cyberattacks

Artificial Intelligence (AI) is the hottest piece of tech that we saw coming, and its swift progress can occasionally be unsettling. Some said we’re advancing way too fast with AI, and with so little time yet so much progress, AI is prone to misuse and vulnerabilities.

Microsoft has recently outlined its approach to tackling these issues with generative AI, particularly large language models (LLMs). The Redmond tech giant is often considered the world’s leader in the generative AI market, thanks to the Copilot AI assistant tool on Windows and mobile. 

Microsoft says that it limits what the LLM can and will do using several layers of defenses through Microsoft AI’s architecture. The layers are meant to stop two main types of attacks on AI systems, whether it’s malicious prompts or poisoned content. 

Microsoft has developed a technique called Spotlighting, which marks external data to differentiate it from legitimate instructions. For instance, a bad email could trick the AI into revealing sensitive information from the user’s emails, and this system can help the AI tell it apart from real commands without slowing it down. 

And it’s extremely helpful, considering how Copilot will soon be everywhere in Outlook, the popular email service. Microsoft is also aware of a new kind of cyberattack called Crescendo that can manipulate AI systems like Copilot into creating harmful content. 

“In the case of Crescendo, our teams made software updates to the LLM technology behind Microsoft’s AI offerings, including our Copilot AI assistants, to mitigate the impact of this multiturn AI guardrail bypass,” Microsoft promises. 

OpenAI, the rising AI company with hefty financial support from Microsoft, has also limited the use of AI for political campaigns, as it attempts to promote healthier, no-misuse of AI.