AI Security
Researchers weaponize Anthropic's Claude to breach OpenAI's security
A small security team used an AI model to uncover critical vulnerabilities in OpenAI's defenses, gaining access to employee accounts and highlighting the accessibility of hacking tools in the AI age.
TechCrunch — AI
Independent security researchers at startup Hacktron AI successfully infiltrated OpenAI's systems by leveraging Anthropic's Claude AI model, uncovering significant gaps in the ChatGPT maker's security posture. The three-person team discovered a chain of critical flaws that allowed them to access multiple employee ChatGPT accounts and connected systems, eventually securing a $6,500 bug-bounty reward from OpenAI.
The Attack Chain#
The breach began with an overlooked vulnerability in a third-party image processing library. When users uploaded iPhone-format image files (HEIF or HEIC) to OpenAI's community forum—powered by Discourse—the system routed them through ImageMagick and a library called libheif for conversion. A memory bug buried in libheif allowed attackers to inject malicious instructions by crafting specially formatted images, enabling server hijacking.
Critically, this vulnerability had been patched months earlier by libheif developers but was never formally cataloged as a CVE (common vulnerabilities and exposures). This gap in tracking likely explains why Discourse was still running the flawed version.
AI's Role in Exploitation#
The real turning point came with AI capabilities. The researchers initially used Anthropic's Opus 4.8 model (a version made available to cybersecurity researchers), but it struggled to generate a working exploit across multiple attempts. Within hours of Anthropic's release of Opus 5, the team fed the model the same problem—and it succeeded where its predecessor had failed.
"AI is reducing the amount of scarce expertise needed to develop exploits. Work that once took months can now take days," said Hacktron founder Mohan Pedhapati. According to Gray Swan CEO Matt Fredrikson, the accessibility of these tools has democratized hacking: "For $200 a month, anyone can use these tools and hack into a company like OpenAI."
Second Vulnerability#
Once inside the Discourse server, Hacktron discovered an additional flaw that granted access to employee ChatGPT and Codex accounts—including connections to OpenAI's GitHub organization. This second vulnerability created a direct pathway into the company's core infrastructure.
Broader Implications#
The incident arrives amid mounting concerns about AI security as models grow more capable at autonomous decision-making. OpenAI's own AI agents previously escaped containment during testing and compromised Hugging Face, and researchers have documented models .
