Multiple ChatGPT Security Bugs Allow Rampant Data Theft

Dark Reading reports that seven security flaws in OpenAI’s ChatGPT and its SearchGPT browsing feature could enable attackers to exfiltrate private user data, bypass safety filters and manipulate the chatbot’s behaviour. Researchers from Tenable describe proofs of concept showing how malicious prompts hiding in blog comments, poisoned search results and exploitative use of URLs can be exploited, with zero-click and one-click attack vectors identified as especially dangerous.

The flaws arise from how ChatGPT ingests external content, stores memories of conversations, and handles web browsing, including responses to URLs and Bing-tracking links. Tenable’s findings, which are described as potentially affecting millions of users, indicate that an attacker could chain multiple vulnerabilities to create an attack path spanning injection, evasion, data exfiltration and persistence.

OpenAI was informed of the disclosures in April, and while the company acknowledged receiving them, it has not publicly confirmed changes. According to Tenable, the researchers emphasise that a high-resource actor could exploit these vulnerabilities to target multiple users, including through comments on product reviews and prompts that link to phishing sites.