A group of hackers breached the OpenAI codebase using Claude 3 Opus

Three security researchers successfully hacked into OpenAI's core codebase using the capabilities of the Claude 3 Opus model. The incident occurred within hours of the model's public release, allowing the attackers to access critical research data and the company's production code. To carry out the attack, the hackers spent less than $3000 on computing power. The use of LLM enabled the automation of the vulnerability discovery process and analysis of complex architectural decisions, which previously required significant time investments from cybersecurity specialists. During the operation, the group members were able to bypass standard protection protocols, exploiting the model's ability to quickly write exploits and interpret closed system data. The access gained covered almost all key repositories, including internal development tools and infrastructure configurations. This case demonstrates new risks associated with the use of advanced generative models for offensive purposes. The attack cost, not exceeding $3000, highlights the affordability of tools for carrying out large-scale cyberdiversions against high-tech corporations. The incident raises questions about current methods of protecting proprietary code from AI agents. Developers have to revise their security approaches, considering that modern models can find critical code errors faster than previous-generation vulnerability scanners.