Not all coding problems are AI-generated. On Tuesday, Anthropic said a leak triggered by a “human error” exposed 500,000 lines of source code underlying its marquee AI coding assistant Claude Code. The company said the leak didn’t impact sensitive customer data or its AI models, but it did reveal some details about how Anthropic built the tool’s user interface. By Wednesday, Anthropic used copyright takedown requests to pull down 8,000 copies posted on GitHub.

It’s a bad look for Anthropic. Claude Code, which brought in $2.5 billion in run rate revenue as of February, has helped it gain an edge against its rivals and boost its business.

Worse, it’s Anthropic’s second data mishap of the week. On Monday, the company accidentally published a blog post announcing its next big model update: Claude Mythos, The Information reported.

Although the Claude Code incident was human error, security breaches aren’t isolated incidents. Data labelling startup Mercor confirmed that it was hit by a security incident linked to open source project LiteLLM, TechCrunch reported. Extortion group Lapsus$ claimed it had gotten access to data like emails, phone numbers and resumes of data labelers as well as source code but it remains unclear if that data was affected by the breach. These types of attacks are bound to be more prevalent as AI agents write code and ship software, Sarah Guo, founder of AI-focused VC firm Conviction wrote in a post on X.

Guo calls it “dark code.” Before AI, when all code was written slowly and deliberately by humans, programmers were forced to deeply understand the systems they were building. But now that AI agents write code super-fast, no one fully understands it or the decisions the agent made, making it difficult to pinpoint why a data leak or security incident happened. Because coding agents select tools and execute plans in real time, documentation of........

© Forbes