939 gigabytes. That's how much source code was exfiltrated from Mercor, the AI-powered hiring platform, through Claude.

The breach wasn't a traditional hack. It exploited the AI agent's access to the company's code repositories — the same access granted to enable Claude to help with development tasks. An attacker leveraged that access path to extract Mercor's entire codebase — nearly a terabyte of proprietary source code.

The scale was staggering. 939GB isn't a few files or a single repository. It's an organization's complete intellectual property — algorithms, infrastructure code, customer-handling logic, internal tools, and every secret embedded in every config file across the entire operation.

The incident demonstrated the catastrophic downside of granting AI agents broad code access: every byte the agent can read is a byte that can be exfiltrated. Access controls that work for human developers — who read code file by file — don't work for AI agents that can systematically traverse and extract entire codebases in minutes.

When you give an AI access to your code, you're not just trusting the AI. You're trusting every possible way that AI can be compromised.