[Two Cache Bugs in Claude Code Exposed, Silently Driving API Costs Up by 10-20x] According to 1M AI News, a developer reverse-engineered the 228MB binary file of the standalone version of Claude Code using Ghidra, MITM proxy, and radare2, uncovering two separate cache bugs that can silently increase API costs by 10-20x without the user's knowledge. The related analysis has been submitted to GitHub (issue 40524) and has been marked by Anthropic as a regression bug, with a team assigned to address it. The first bug exists in the custom Bun runtime used by the standalone version, where an error in the billing identifier replacement logic causes a full cache rebuild to be triggered with every request. The second bug affects users who restore sessions using the --resume or --continue flags. Introduced in v2.1.69, this bug causes a cache prefix mismatch, leading to the entire conversation history being rewritten instead of being read from the cache. Previously, Anthropic engineers had confirmed that users were reaching usage limits "much faster than expected."