Known Bugs and Regressions

Documented issues in claude-code discovered through community analysis and the source leak.

CVE Cluster

CVE CVSS Mechanism Fixed
CVE-2025-59536 8.7 Pre-trust hook execution — RCE before trust dialog v1.0.111
CVE-2026-21852 5.3 ANTHROPIC_BASE_URL redirect exfiltrates API key v2.0.65
CVE-2025-54795 8.7 Confirmation prompt bypass via prompt crafting v1.0.20
Adversa deny-cap TBD 50-subcommand overflow in bash-security v2.1.90
CVE-2026-35020/21/22 TBD Command injection in which.ts, promptEditor.ts, auth.ts Unpatched

Cache Bug Catalogue (B1-B6)

Bug Description Status
B1-B3 Cache prefix invalidation bugs Fixed in v2.1.91
B4 Synthetic entry injection — 151+ fake entries from background tasks Open
B5 Cache alignment bug Fixed in v2.1.91
B6 Zero-reasoning budget — adaptive thinking set high but individual turns get 0 tokens Fixed in v2.1.94

v2.1.91 was the most important stability release. v2.1.94 changed default effort from medium to high.

Performance Issues

Quality Regression (Feb-March 2026)

Quantified by Stella Laurenzo investigation (17,871 thinking blocks, 234,760 tool calls): - Median thinking depth collapsed 67-75% (2,200 → 560-720 chars) - Read:Edit ratio collapsed from 6.6:1 to 2.0:1 - Full rewrites doubled (4.9% → 10-11%) - User vocabulary shifted: "simplest" +642%, "stop" +87%, "great" -47% - Cost exploded 122x despite same user prompt volume

Three Streaming Hang Bugs

Cascading watchdog bugs: (1) armed after dangerous phase, (2) targets undefined objects, (3) fallback fires in wrong phase. Kolkov analysis: 16.3% API failure rate, 5.4% orphaned tool calls.

cch=00000 Bug

native-attestation sentinel string replacement corrupts message content when discussing billing, causing 10-20x token consumption per request.