Yesterday, on March 31, 2026, Anthropic accidentally "open‑sourced” its Claude Code codebase. A packaging mistake in @anthropic‑ai/claude‑code v2.1.88 left a ~60 MB source‑map file on npm, exposing about 512,000 lines of internal TypeScript across ~1,900 files.
Within hours, the internet had:
- 44 hidden feature flags
- Unreleased agents, cron jobs, and browser automation via Playwright
- Internal prompts and early hints of a "Mythos" / "Capybara"‑style agent model
Anthropic spokesperson said "no customer data was exposed" , calling it a "release packaging issue caused by human error", not a breach.
What this leak exposes
- Build artifacts are now crown jewels
This wasn’t a hack; it was a misconfigured build. Source‑maps, debug files, and internal tooling are now first‑class attack surfaces. If they ship in public packages, the whole codebase can be reconstructed. - Roadmaps leak in minutes
Hidden flags and internal prompts effectively map out Anthropic’s roadmap. One wrong config, and your “next‑gen” features become public demos before launch. - Trust in AI‑native companies is fragile
Anthropic brands itself as “careful,” yet this is at least the second major leak‑style incident in a short span. When AI‑native players ship dangerous build artifacts, it erodes confidence that they’re inherently more secure.
What this means for your software strategy
At Dodera Software we design systems that are secure, scalable, and resilient by default.
This incident reinforces three principles we already bake into our workflows:
- Every build is inspected, not assumed
- We strip debug info and validate every artifact before it ships.
- Source‑maps, internal configs, and prototypes never make it into public packages by accident.
- Security is part of the AI stack
- When we integrate Claude Code, Copilot, or other AI agents, we control how they access code, data, and infrastructure.
- Roles, permissions, and audit trails are baked into AI‑driven workflows.
- Roadmaps are guarded, even in code
- Experimental features and internal prototypes live in isolated branches or repos.
- Feature flags are coupled with access controls so leaks don’t instantly expose your next big move.
How to future‑proof your AI‑native stack
If you’re building or planning AI‑native products, tighten this today:
- Audit your build outputs
Check what actually ends up in your npm packages, Docker images, or deployment bundles. If it wasn’t meant for the public eye, it shouldn’t be there. - Treat source‑maps as sensitive
Never ship unstripped source‑maps in production or public packages without explicit, intentional design. - Isolate internal prototypes
Keep future‑looking features in private repos or behind strict access controls. - Design for incident‑friendly code
Assume some part of your stack might leak. What’s the impact if internal prompts, agent logic, or roadmap features go public?
The takeaway for our clients
This leak isn’t just a "big‑tech" story. It’s a live example of how fast reputation, IP, and competitive advantage can slip away when one configuration step goes wrong.
At Dodera, we help you:
- Build secure, production‑ready AI workflows.
- Integrate AI agents into your stack with clear guardrails and audit trails.
- Design future‑proof architectures that assume leaks can happen and minimize their impact.
If you’re using or planning to use AI‑native tools like Claude Code, GitHub Copilot, or custom agents, now is the time to review how they’re packaged, published, and secured.