Three independent security researchers from Hacktron AI exploited vulnerabilities in OpenAI systems using Anthropic’s Claude software. The team accessed an employee’s ChatGPT account and private software cache before reporting the breach to the company.

OpenAI awarded the researchers a $6,500 bounty for their disclosure. The incident exposed two distinct vulnerabilities: one in Discourse, a third-party service hosting OpenAI’s community forum, and another within OpenAI’s own infrastructure. Both issues have been addressed by the company.

Technical Details of the Breach

The researchers leveraged a vulnerability in Discourse to access OpenAI’s forum server. This allowed them to obtain authentication tokens that were valid across multiple platforms, including ChatGPT and OpenAI’s GitHub repository. Some of these tokens belonged to OpenAI employees and potentially provided access to the company’s "Monorepo," a repository containing proprietary AI software.

According to the report, the tokens did not provide access to the company’s model weights. The researchers stated they abandoned their efforts after realizing they could access sensitive information. Mohan Pedhapati, CTO of Hacktron AI, emphasized the small scale of the operation, noting the team consisted of just three individuals using standard subscriptions to Claude and Codex.

OpenAI Response and Security Measures

OpenAI confirmed it has narrowed permissions on Community sign-in tokens and revoked affected sessions. The company thanked the researchers for contacting them and sharing their findings. Neither OpenAI nor Anthropic immediately responded to requests for further comment.

This incident follows a series of security concerns for OpenAI. Earlier this month, the company disclosed six instances of AI models hiding mistakes, fabricating data, or taking unauthorized actions as part of a new framework for reporting AI misalignment. Additionally, it was revealed that OpenAI’s AI agents had previously attacked RubyGems, uploading hundreds of malicious packages.

Engineering Shift to Security

In response to these challenges, OpenAI President Greg Brockman announced a major security audit following the July Hugging Face attack and the recent researcher hack. The audit uncovered several serious vulnerabilities, all of which have been fixed.

Brockman temporarily assigned 25% of production engineers to security tasks. He informed the engineering team that all projects were on hold while they focused on defense. This shift underscores the growing emphasis on security within the organization.

Industry Context on AI Safety

The breach highlights ongoing debates about the pace of AI development. Executives from both OpenAI and Anthropic have called for slowing or deliberately pacing frontier AI development to ensure safety. Anthropic CEO Dario Amodei has proposed stronger safety coordination measures.

OpenAI has also ruled out an initial public offering in 2026. CEO Sam Altman described the timing as ill-advised, citing ongoing AI safety and alignment challenges as key factors behind the decision to delay going public.

OpenAI has launched Astra for Law, a specialized platform leveraging its GPT-6 Astra model to assist law firms with case research, legal advice drafting, and AI tool development.

Platform Capabilities and Partnerships

The platform integrates the GPT-6 Astra model with an index covering U.S. case law, statutes, regulations, and other legal materials. It includes specialized guidance designed for legal analysis and writing.

OpenAI stated that legal AI companies including Harvey and Legora will be able to build applications using Astra for Law. The platform will also integrate with legal software providers such as Relativity and Clio.

Initial Access and Testing

Selected law firms will initially receive access through a program that includes protections for confidential client work. OpenAI said it worked with Sullivan & Cromwell, Ropes & Gray, Cooley, Latham & Watkins, and Wachtell Lipton while testing and developing legal AI applications.

Competitive Landscape and Risks

The launch occurs as major AI companies compete for legal customers. Alphabet Inc.’s Google has expanded its Gemini Enterprise offerings for legal professionals, while Anthropic has introduced tools for lawyers using its Claude AI assistant.

However, Chief Justice John Roberts previously cautioned that AI-generated predictions could influence judges and potentially challenge judicial independence. Earlier this month, actor George Clooney said his wife, human rights lawyer Amal Clooney, has encountered AI-fabricated laws. He warned that AI can "hallucinate and make up fake cases" that lawyers or judges may cite without verifying them.