On March 31, a security researcher found Anthropic's Claude Code source code sitting in a public npm package. A forgotten source map exposed 512,000 lines of raw TypeScript, 44 unreleased feature flags, and internal model codenames. Anyone could download it.
This is the second time Anthropic leaked source code through the same mechanism in 13 months.
The embarrassment is one thing. What matters more is what the leaked code tells us about Anthropic's roadmap, and why SaaS and automation startups should be paying very close attention.
How Claude Code's source code leaked via npm
During the v2.1.88 npm release, someone at Anthropic missed a .npmignore configuration. That pushed a 59.8 MB source map file to the public npm registry, which also contained a link to a Cloudflare R2 bucket with the full, uncompressed TypeScript archive.
Chaofan Shou, a UC Berkeley PhD candidate, spotted it and posted the download link on X. Within hours, mirrored GitHub repos racked up 84,000+ stars before Anthropic filed DMCA takedowns. Anthropic called it a "release packaging issue caused by human error" and confirmed no customer data was exposed.
What Claude Code's leaked feature flags reveal
The 44 hidden feature flags sketch out something much bigger than a coding tool. Browser automation (Playwright-based). Scheduled triggers where agents self-schedule tasks via cron, the same thing Zapier, n8n, and Make.com sell today. Multi-agent orchestration. Voice commands. Cross-conversation communication via Unix sockets. Team memory sync.
Less like a coding assistant roadmap, more like the spec for an AI operating system.
Why KAIROS could sunset most automation and SaaS tools
One feature stands out. Codenamed KAIROS, it appears over 150 times in the source code. It turns Claude Code from a "respond when asked" tool into a daemon that runs continuously in the background, monitors your repos, subscribes to GitHub webhooks, pushes notifications to your phone, and suggests code changes while you sleep.
KAIROS has its own toolset: file sending, push notifications, PR monitoring, background pull request suggestions. It even includes a memory consolidation engine called autoDream that runs through phases called Orient, Collect, Consolidate, and Prune. The system prompt literally reads: "You are executing a dream."
This is a direct threat to every SaaS tool that automates workflows, schedules tasks, or connects APIs. And not just Anthropic. OpenAI's Codex, Google's Jules, and Gemini CLI are all moving in the same direction. The SaaSpocalypse we wrote about is getting very real.
What this means for Malaysian SaaS and automation startups
If you're building an AI wrapper, a workflow automation tool, or a thin SaaS layer on top of an LLM API in Malaysia, this leaked roadmap is a wake-up call.
The features KAIROS bundles, scheduled triggers, browser control, background execution, webhook subscriptions, are what dozens of startups sell as standalone products. When the platform ships these for free, your value proposition disappears overnight. Ira Bodnar, founder of AI ad company Ryze, saw her close rate drop from 70% to 20% after a single Claude feature release. That's how fast it happens.
The businesses that survive platform absorption are the ones building where AI companies structurally won't go:
- Regulated verticals. Healthcare (FDA, HIPAA), finance (SEC audits), legal compliance. None of the 44 leaked feature flags touch domain-specific regulation. PDPA compliance in Malaysia is exactly this kind of moat.
- Deep system integration. Healthcare requires integrating 40+ medical systems. Construction needs field-level adoption. Legacy system work means reverse-engineering APIs that don't exist. AI companies have no incentive to do this.
- Outcome-based pricing. AI platforms charge per token. Vertical AI companies charge per result: per invoice processed, per ticket resolved. That competes for labor budgets, not IT budgets.
If your product is a thin wrapper, go deeper now. Add industry-specific workflows and proprietary data flywheels that a general-purpose platform can't replicate.
Our take
We use Claude Code every day at Gotchaa Lab. KAIROS and the rest of this roadmap don't scare us, because we don't sell thin wrappers. We sell judgment: architecture decisions, Malaysian business context, PDPA compliance, and accountability when things break. Those are exactly the things AI platforms can't and won't absorb.
But if your entire business model is "connect API A to API B with a nice dashboard," this leak just showed you the expiry date. The platform is coming for that layer. The smart move is to go deeper into your vertical before it arrives.
Talk to us if you want an honest take on where AI fits in your business.
