How Three Indian-Origin Researchers Hacked OpenAI With Claude — and Got Paid $6,500

In July 2026, three Indian-origin cybersecurity researchers at the US-based startup Hacktron AI breached OpenAI's systems — with permission — using Anthropic's Claude chatbot to help write the exploit code. The team, Harsh Jaiswal, Mohan Pedhapati and Rahul Maini, moved from OpenAI's public community forum to employee ChatGPT accounts and an internal GitHub repository in under 72 hours, spending less than $3,000 on AI tokens. They reported the flaws responsibly, and OpenAI patched them and paid a $6,500 bug bounty, according to reports published on 18 September 2026 by The Guardian and the Times of India.

How the OpenAI hack worked, step by step

The attack began at community.openai.com, OpenAI's community and help forum, which runs on the third-party Discourse platform. The Hacktron team found that specially crafted HEIC/HEIF image files could exploit a flaw in the libheif image-decoding library used by Discourse, allowing remote code execution on the forum server.

From there, they chained a second vulnerability to pivot into OpenAI employee ChatGPT and Codex accounts. To prove access without touching sensitive intellectual property, they used a compromised employee's Codex account to submit a harmless "pull request" to an internal private GitHub repository. The researchers stressed they had access to the code but did not download or view it.

The role of Claude — and OpenAI's own model

Claude was used initially to generate code for the intrusion, but the researchers said they largely relied on OpenAI's own cutting-edge GPT-5.6 Sol model to carry out much of the hack — a detail first reported by the Wall Street Journal. The irony was not lost on observers: one AI lab's models were used to break into another's, and the target's own tools helped finish the job.

What OpenAI did next

The findings were disclosed through OpenAI's bug bounty channels, including Bugcrowd and HackerOne, and also reported to Discourse. OpenAI patched the identity issue shortly after notification. A company spokesperson thanked the researchers for sharing their findings and confirmed the exploited vulnerabilities had been addressed. There is no indication of any ongoing threat; this was an authorised test under OpenAI's ethical hacking programme, not a criminal breach.

Why this matters for AI security

The episode illustrates a shift cybersecurity experts have been warning about: AI tools compress work that once took a well-resourced team months into a matter of days, for a few thousand dollars in tokens. It also fits a pattern — OpenAI had earlier revealed that a "swarm" of AI agents powered by its technology hacked the startup Hugging Face during a security test. The practical lessons for AI companies and enterprises:

• Audit third-party components — here, a forum platform and an image library were the entry point, not OpenAI's core systems.
• Treat employee accounts and internal repositories as high-value targets reachable through chained, low-severity bugs.
• Run and fund bug bounty programmes; this breach cost OpenAI $6,500 instead of a potentially far more damaging malicious disclosure.

Tech commentator Deedy Das noted that none of the three researchers had brand-name colleges or companies on their CVs — "just raw curiosity and skill." For Indian readers, the story is also a reminder that the country's cybersecurity talent is operating at the frontier of AI safety research globally.