OpenAI Tightens Access After Vulnerability Reported by Hacktron

OpenAI has revoked tokens and restricted permissions following the discovery of a vulnerability affecting its community forum. Hacktron claims to have exploited the flaw using Claude as part of an Anthropic program, ultimately compromising an employee account and prompting Codex to suggest modifications, while stopping short of any access to internal code. The startup states it received $6,500 and claims to have fewer than 10 employees.

OpenAI Revokes Tokens and Hacktron Claims to Have Stopped Before Accessing Code

OpenAI indicated that it has restricted permissions on community login tokens and revoked the affected tokens and sessions after the vulnerability was discovered. Hacktron clarifies that its team stopped before accessing OpenAI's internal code. The startup reports having received a $6,500 bounty for this discovery.

A Community Forum Vulnerability Exploited via Claude

In July, the Hacktron team identified vulnerabilities within OpenAI's infrastructure. According to a disclosure made public on a Sunday, Hacktron states that a user or employee's login to OpenAI's community help forum exposed their ChatGPT and Codex accounts to a risk of compromise. The team used Claude to attempt to exploit this vulnerability, taking advantage of Anthropic's Cyber Verification Program, which, according to Zayne Zhang, alleviates certain limitations for authorized security research. Hacktron successfully compromised an OpenAI employee's account and prompted that employee's Codex account to suggest modifications in the internal code repository before reporting the issue to OpenAI.

Who is Hacktron and What Context Surrounds the Incident

Hacktron is a cybersecurity startup based in San Francisco, co-founded and led by Zayne Zhang. According to Zhang, the Hacktron team began auditing leading AI companies like OpenAI to identify potential exploitable vulnerabilities by AI agents. The company was founded less than a year ago and has fewer than 10 employees. Zhang believes that AI security and cybersecurity are converging and wants to see more cybersecurity experts involved in the sector. Anthropic did not respond to a request for comment. In recent months, OpenAI, Anthropic, and Meta have revealed that their agents conducted unauthorized actions during tests, as fears of apocalyptic scenarios related to uncontrolled malicious AI agents have emerged this month. The incident comes shortly after a hacking incident involving OpenAI on Hugging Face. Hacktron reports having used Claude to target OpenAI's internal code repository and has reported vulnerabilities.