Three Researchers Used Claude to Access OpenAI's Internal Code
Three researchers at Hacktron AI utilized Anthropic’s Claude to exploit two existing web vulnerabilities and gain access to OpenAI employee accounts and an internal code repository within 72 hours. They disclosed the issue privately, and OpenAI patched the vulnerability in approximately 14 hours, for a reward of $6,500.
Neither of these weaknesses was related to AI; they were common web security flaws found in third-party software. The research paper citing this incident states that its authors are now less confident in the industry's progress.
Security Researchers and Chain Exploits
Security researchers Harsh Jaiswal, Mohan Pedhapati, and Rahul Maini successfully navigated these vulnerabilities to access OpenAI's internal systems. This sequence is often referred to as a "bug bounty" – a scenario where researchers responsibly disclose security issues to vendors for remediation and compensation.
The Role of AI in Security
While the framing of this event portrays AI as a tool for breaching security, it's important to note that:
- The vulnerabilities were not AI-related. They stemmed from issues in forum software and session token management, both well-known web security problems.
- AI accelerated the exploitation process, but the underlying flaws are not new.
The Impact of Patch Windows
The significant reduction in time between discovery and exploitation of these vulnerabilities highlights a broader concern about patch windows. As AI models improve, defenders must ensure their tools keep pace to address emerging security challenges effectively.
Payment for Access
The $6,500 reward for this chain of exploits raises questions about the value of such access. Considering OpenAI's massive valuation, this payment might seem disproportionate, but it aligns with a pattern TNW has observed:
- Anthropic, Google, and Microsoft have offered similar bounties for agent vulnerabilities without public disclosure.
Bounty programs offer valuable feedback on the perceived severity of specific security issues within the industry.