Microsoft revealed a new security vulnerability called “Whisper Leak,” which allows attackers to identify conversation topics in chatbots based on artificial intelligence techniques, by analyzing data sent over the network without the need for decryption.
The company explained, in a report published on November 7, that the vulnerability allows attackers to collect sensitive information about the nature of instructions provided to smart systems, by monitoring the size of packets and their transmission speed. The tests demonstrated that the attacker was able to classify targeted conversations with up to 100% accuracy, and to capture between 5% and 50% of conversations related to one topic.
Microsoft warned of the effects of this vulnerability on users’ privacy, especially in environments with tight Internet censorship, where talking about topics such as protests, elections, or sensitive files may expose users to monitoring or tracking. She also pointed out that entities such as Internet service providers or some government agencies may be able to identify users who discuss specific issues, such as money laundering or political opposition.
In a rapid response, several technology companies, including OpenAI, Mistral, Microsoft, and AIX, announced the implementation of new protection measures.
The measures adopted by “OpenAI” and “Microsoft Azure” included adding random texts of varying length to the chatbots’ responses, with the aim of hiding the original structure of the programming sentences and preventing them from being traced.
Microsoft also recommended that users avoid highly sensitive discussions over untrusted networks, be sure to use virtual private networks (VPN), and follow the security guidelines announced by companies operating digital services.
It should be noted that this disclosure comes after the announcement of another group of vulnerabilities that may affect conversation models such as “GPT Chat,” which may allow attackers to access personal information from user records and conversation memory.
