Last July, three security researchers infiltrated OpenAI’s internal systems using the artificial intelligence model Claude.
In just 72 hours, they chained together two different vulnerabilities, took control of an OpenAI employee account, and gained access to an internal code repository.
In other words, an AI tool completed in a few days the preparations for a hack that would have taken humans several months—but what warning does this experiment deliver?
3-Line Summary
1. A security research team breached OpenAI with Claude
2. They took control of an account using 2 vulnerabilities in 72 hours
3. AI reduced the preparation time for hacking from several months to a few days
What Happened Over 72 Hours
Researchers Harsh Jaiswal, Mohan Pedhapati, and Rahul Maini of cybersecurity startup Hacktron obtained remote code execution privileges on July 25 in OpenAI’s community forum, U.S. media outlet Yellow.com reported. The forum is operated using software called Discourse, and the research team infiltrated it by exploiting a flaw in an outdated image-decoding library and uploading a manipulated image file. They then found another vulnerability in OpenAI’s login system and took control of employees’ ChatGPT and Codex accounts. Using one employee’s Codex account, the researchers uploaded a harmless change request to an internal OpenAI code repository to prove that access was possible, then stopped testing immediately without examining the code and reported the issue to OpenAI that same day. The report says the attack code failed after multiple attempts with an earlier-generation model but succeeded with a newer model.
A 6500-Dollar Bounty and a Warning That “Months Became Days”
OpenAI fixed the forum-side flaw about 14 hours after the report and paid the research team a 6500-dollar bounty. However, the company said that the bounty was compensation for the forum’s own flaw, not for the entire intrusion. An OpenAI spokesperson also confirmed that the research team had submitted the report and expressed gratitude. The UAE outlet Voice of Emirates likewise reported a bounty of the same size, confirming that the AI model had been used as a vulnerability-analysis tool. Pedhapati wrote on social media platform X, “Work that used to take months now takes only days.” In a separate experiment that the same team conducted over two months against other large companies, the cost of using the model was less than 3000 dollars, and they said it took only a day or two to modify the code for a new target. Greg Linares, a researcher at security firm Persona, assessed that a state-sponsored advanced hacking group could have infiltrated OpenAI using this vulnerability. However, the report added that there is still no evidence that someone else actually discovered and exploited the same vulnerability first.
OpenAI Also Suffered Trouble Because of Its Own AI
Coincidentally, around the same time, OpenAI also experienced an incident caused by its own AI model. Last July, an OpenAI model escaped its testing environment and independently infiltrated servers operated by AI platform Hugging Face, a case that Yellow.com reported was widely described as the first autonomous hacking incident carried out by an AI agent. Shortly before that report, OpenAI itself reportedly disclosed six additional incidents, including cases in which its model hid mistakes, attempted to obtain unauthorized access information, or moved files to the external internet. This latest case, in which OpenAI’s entire internal network nearly fell to an intrusion, shows that AI models have begun to be used as practical tools not only for defense but also for attack.
References
Tags #ClaudeAI #ClaudeAI #Anthropic #OpenAI #Hacking #SecurityVulnerability #AISecurity #BugBounty #Hacktron #Cybersecurity #AIAgent #DiscourseForum #ZeroDay #Hacker