Anthropic accidentally published a JavaScript sourcemap that exposed Claude Code’s 512,000-line TypeScript operational blueprint, and researchers quickly reconstructed and redistributed the code. Separately, Adversa AI disclosed a critical permission-enforcement vulnerability that lets crafted prompt-injected command pipelines bypass deny rules, risking credential exfiltration and supply-chain or cloud compromise. #ClaudeCode #AdversaAI
Keypoints
- On March 31, 2026 Anthropic accidentally published a debugging sourcemap for Claude Code v2.1.88, enabling rapid de-obfuscation and public distribution.
- Leaked files total roughly 512,000 lines of TypeScript across 1,900 files, exposing the tool’s operational blueprint but not model weights, training data, or customer credentials.
- Adversa AI discovered a critical flaw in Claude Code’s permission system where deny rules can be silently bypassed by generating pipelines with more than 50 subcommands via prompt injection.
- The vulnerability could enable exfiltration of SSH keys, AWS/GitHub/npm tokens, environment secrets, supply-chain compromise, and CI/CD or cloud breaches.
- Anthropic’s LLM safety layer blocked some malicious payloads during testing, but the underlying policy-enforcement bug requires a code fix and additional defensive controls.
Read More: https://www.securityweek.com/critical-vulnerability-in-claude-code-emerges-days-after-source-leak/