Claude Code source code exposure and reported performance degradation
Anthropic's Claude Code npm package exposed unredacted source code (46,000+ lines) including core reasoning modules when a configuration error left a source map file publicly accessible on 31 March 2026. Separately, AMD's AI director reported analysis showing Claude Code exhibited degraded reasoning and increased lazy behaviors following deployment of a thinking redaction feature in early March.
What happened
On 31 March 2026, Anthropic's npm package distribution of Claude Code contained an unredacted 'cli.js.map' source map file (59.8 MB), which allowed restoration of 4,756 source files including the core QueryEngine.ts reasoning module and 1,800+ core source code files with developer comments. A security researcher named Chaofan Shou discovered the file, and the exposure was widely distributed through GitHub backups.
Beginning in early March 2026, AMD AI director Stella Laurenzo reported that Claude Code exhibited degraded reasoning and performance. Her analysis of 6,852 Claude Code sessions spanning 234,760 tool calls and 17,871 reasoning blocks showed a spike in 'lazy' behaviors (evasion of responsibility, early stopping, frequent permission requests) from zero instances before 8 March to an average of 10 per day by late March. Code reading frequency before modification dropped from an average of 6.6 times to 2 times by late March, and the system began rewriting entire files instead of making targeted modifications.
Reported impact
Laurenzo attributed the performance decline to the 'thinking content redaction' feature deployed in Claude Code version 2.1.69 in early March, which by default removes reasoning content from API responses, preventing users from understanding the model's internal reasoning. She stated that shallow thinking causes the model to take lowest-cost actions: modifying without reading, stopping prematurely, avoiding responsibility for errors, and choosing simplest rather than correct solutions.
The source code exposure revealed previously undisclosed features including 'Kairos' (an autonomous background daemon with session persistence) and 'ULTRAPLAN' (a mode offloading complex planning to remote cloud containers with 30-minute Opus 4.6 reasoning). Laurenzo stated her team switched to alternative vendors with better product quality, and that the exposure could alter industry competitive dynamics by reducing the engineering moat of Anthropic's products.
Index analysis
The source map leak exposed Claude Code's engineering architecture rather than model weights or user data. The exposure included 4.6 million characters of core reasoning code and developer implementation notes, which competitors could use to replicate Anthropic's agent design patterns and optimization logic without replicating the underlying model.
The incident follows a similar source map exposure in February 2025 involving an earlier Claude Code version. Laurenzo requested that Anthropic disclose whether thinking token allocation was restricted, provide per-request thinking token metrics, and offer differentiated subscription tiers for users requiring deep reasoning (up to 20,000 tokens) versus lighter use cases (200 tokens).
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.