Microsoft Corp. Bolsters Security Features in Azure AI Studio to Combat Malicious Attacks
In an effort to enhance security measures and prevent malicious attacks on its AI chatbots, Microsoft Corp. has announced new tools and features in Azure AI Studio. The company is introducing prompt shields to detect and block prompt injection attacks and indirect prompt injections that can trick AI chatbots into behaving in unintended ways.
Sarah Bird, Microsoft’s chief product officer of responsible AI, emphasized the unique challenge and threat posed by these types of attacks. The new defenses are designed to identify suspicious inputs in real time and prevent unauthorized actions, such as stealing user information or hijacking systems.
One of the new features being implemented by Microsoft is the ability to alert users when a model generates false responses. This move comes in light of incidents involving its Copilot chatbot generating harmful responses, leading the company to prioritize building trust in its generative AI tools.
According to Bird, tell-tale signs of attacks include asking a chatbot repetitive questions or role-playing prompts. Microsoft, as the largest investor in OpenAI, is committed to the safe deployment of AI and is building protections into its large language models.
Bird further emphasized that relying solely on the model technology is not enough, as jailbreaks remain an inherent weakness. By implementing these new security features, Microsoft aims to safeguard its AI chatbots and improve trust in the technology.
“Infuriatingly humble tv expert. Friendly student. Travel fanatic. Bacon fan. Unable to type with boxing gloves on.”