Video: "New Claude Sonnet 4.8 LEAKS!" by Julian Goldie on YouTube.

How the leak happened

In late March 2026, a routine release of Anthropic's Claude Code CLI accidentally included a 59.8 MB npm source map file — the kind of file that maps minified JavaScript back to the original TypeScript. Instead of being stripped from the package, it shipped. Anyone who downloaded version 2.1.88 of @anthropic-ai/claude-code had access to roughly 512,000 lines of source code spread across 1,900 files.

Anthropic fixed it quickly, but the contents had already been indexed. It wasn't a malicious breach — just a packaging oversight — but the code exposed a fair amount about where Claude Code is heading.

Claude Sonnet 4.8: what we know

The source code referenced model names directly, confirming that the next release in the Sonnet line is 4.8. Anthropic typically ships a new Sonnet model one to four weeks after the corresponding Opus release. Given that Opus 4.7 is now available, Sonnet 4.8 is expected sometime in May 2026 — which is now.

The leak doesn't detail specific benchmark improvements for 4.8, but the pattern holds: Sonnet models run faster and cost less per token than Opus, while maintaining most of the capability for day-to-day coding tasks. For anyone using Claude Code on real projects, a fresh Sonnet release usually means better handling of longer codebases and improved instruction-following on complex refactors.

KAIROS: the always-on background agent

The more interesting find is KAIROS. The source code describes it as a daemon mode for Claude Code — an always-on background process that monitors your development environment, runs scheduled checks, sends notifications, and continues work without you actively prompting it.

The name comes from ancient Greek, referring to the right moment to act. In code terms, it means Claude Code would watch your project, notice when something needs attention — a failing test, a PR that needs review, a build that's broken — and act on it. The references in the source also mention a process called autoDream, which appears to be a background task-planning loop that runs when you're not actively using the tool.

To be fair: this is unshipped. Anthropic hasn't announced it and may be some distance from releasing it. But the code is there, not just in notes — it's implemented. That's a meaningful signal.

Undercover Mode and Mythos

Two other references appeared in the source. Undercover Mode seems to describe a Claude Code state where the tool behaves more conservatively — less visible output, reduced logging, designed for situations where you don't want the agent's activity to be obvious. The use case isn't spelled out clearly.

Mythos appears to be a codename, possibly for a next-generation model series sitting above the current Opus/Sonnet/Haiku naming. There's not enough detail to say more than that, and it could refer to an internal project that never ships publicly.

What's worth taking seriously here

Sonnet 4.8 arriving this month is a solid inference, not a rumour. KAIROS existing in the codebase is real, though timing is unknown. The rest is speculation based on code comments and variable names. Worth knowing: model leaks from source code have a decent track record — the same method surfaced Opus 4.7 before its announcement and that proved accurate.

For developers using Claude Code day-to-day, the practical impact is: expect a model update soon, and expect that model to handle longer sessions and bigger codebases more reliably than the current Sonnet. KAIROS, when it ships, would represent a genuine shift — moving Claude Code from something you invoke to something that runs continuously alongside your work.

Where this connects to NordSys

Tracking Claude Code releases — model updates, configuration changes, new features like KAIROS when they eventually ship — is exactly what our AI Agents care plan is built for. If you use Claude Code in your work and want it properly set up and kept current without watching every Anthropic announcement yourself, that's what our AI Agents service handles.

See our AI Agents service →