Claude Code Source Leak Hints at ‘Proactive’ Mode and Experimental Features at Anthropic

A routine update to Anthropic’s Claude Code tool has unexpectedly exposed a large portion of its internals, offering a rare look at experimental features the company may be preparing including a “Proactive” mode and crypto-enabled payments for AI agents.

The issue surfaced after Anthropic pushed version 2.1.88 of Claude Code earlier this week. According to reports from The Verge and others, the release accidentally included a file that revealed the app’s source code. Before Anthropic could correct the error, the codebase was uploaded to a public GitHub repository and rapidly replicated, being copied more than 50,000 times.

In total, more than 512,000 lines of code and around 2,000 TypeScript files were briefly exposed, giving both the broader internet and Anthropic’s competitors a window into how Claude Code is built and what might be coming next.

Once the code was out in the open, developers and enthusiasts began combing through the files for clues. Some of those digging through the repository say they found flags and references to features that are not yet publicly available.

On X, Alex Finn, founder of AI startup Creator Buddy, said he identified a flag for a feature called “Proactive mode.” Based on the code he highlighted, this mode appears intended to let Claude Code operate even when a user has not explicitly prompted it, suggesting more background or autonomous behaviour inside the coding environment.

Finn also claimed to have spotted indications of a crypto-based payment system within the code. The references he pointed to suggest Anthropic has at least explored mechanisms that could allow AI agents to make autonomous payments using cryptocurrency. The exact design, scope or intended use of such a system is not clear from the available description.

Separately, in a Reddit post, another individual reported finding what looked like an experimental virtual companion buried in the codebase described as a Tamagotchi-style character that “reacts to your coding.” According to that post, the feature appeared to be framed as an April Fools-style joke rather than a core product capability.

None of these features have been officially announced by Anthropic, and the company has not publicly detailed them beyond what can be inferred from the code references others have described.

Anthropic has acknowledged the incident, emphasising that it was a packaging mistake rather than a security intrusion.

“A Claude Code release included some internal source code. No sensitive customer data or credentials were involved or exposed,” an Anthropic spokesperson told BleepingComputer. “This was a release packaging issue caused by human error, not a security breach. We’re rolling out measures to prevent this from happening again.”

The company has not shared technical details on those new safeguards, but the statement makes clear that Anthropic sees the incident as an internal process failure rather than an external attack.

While source code visibility can be sensitive from a competitive standpoint, the confirmation that no customer data or credentials were included will be a key concern for users integrating Claude Code into their workflows.

As with many software leaks, there is no guarantee that the capabilities spotted in the code will ever make it into a public release. The presence of flags, configuration options or partial implementations often reflects experiments, internal tools or abandoned directions.

The original report stresses this point: just because a company has written code to support a feature does not mean that feature will ship. Product plans can change, tests can be scrapped and prototypes can remain permanently behind the scenes.

For now, the Claude Code 2.1.88 leak offers a snapshot of what Anthropic’s engineers have at least considered: a more proactive coding assistant that can act without explicit prompts, mechanisms for crypto-based autonomous payments by AI agents and even playful, non-core ideas like a Tamagotchi-style coding companion.

Anthropic says it is putting new measures in place to avoid similar packaging mishaps, but much of the exposed code has already been widely mirrored. How, if at all, this shapes competitors’ strategies or Anthropic’s public roadmap remains to be seen.