CVE-2025-31491

CVE-2025-31491 is a high-severity vulnerability affecting AutoGPT versions prior to 0.6.1, specifically the 'agpt autogpt_platform' product. It allows for the leakage of cross-domain cookies and protected headers, such as Authorization and Proxy-Authorization, during redirect requests. The vulnerability stems from a flaw in AutoGPT's custom requests wrapper, which manually re-requests redirected URLs without properly sanitizing security-sensitive headers for cross-origin requests. The vulnerability has a CVSS score of 8.6 (High), indicating a network-based attack vector with low attack complexity and high confidentiality impact, but no integrity or availability impact. An attacker could exploit this by coercing an AutoGPT instance to visit a malicious open redirect, leading to the unauthorized disclosure of sensitive authentication credentials. Currently, there is no evidence of active exploitation, nor is exploit code publicly available in Metasploit, Nuclei, or ExploitDB. Community discussion and media coverage for this CVE are minimal, which is typical for a significant percentage of reported vulnerabilities.