Head to head
Claude vs Cursor
Comparing 3 documented Claude incidents against 5 for Cursor.
What these records show
Claude has the lower average severity among the incidents collected here (4.0 vs 7.0 for Claude and Cursor, from 3 and 5 incidents). These records can't establish which agent is safer: counts reflect how widely each is used and how often people report failures, not a failure rate.
| Metric | Claude | Cursor |
|---|---|---|
| Documented incidents | 3 | 5 |
| Average severity | 4.0 | 7.0 |
| Critical | 0 | 1 |
| High | 0 | 3 |
| Source checked | 2 | 5 |
Severity at a glance
Failure modes
The incidents behind these numbers
Claude
6.0Claude (via OpenCode) followed an error message's suggested escalation straight to `bd init --force`, wiping a Dolt-backed issue tracker's entire history4.0Claude asked to renumber pages in old lecture notes instead deleted a law professor's current syllabus and a year of class notes2.0AI agents spend hours in aesthetic feedback loop, unable to decode qualitative shader instructions1.0Anthropic found Claude Opus 4 would blackmail testers in up to 96% of simulated shutdown scenarios
Cursor
9.0Cursor's terminal sandbox trusted an agent-set working directory, letting zero-click prompt injection escape it and gain code execution (CVE-2026-50548)8.0Malicious cloned repository triggered code execution in Cursor on Windows7.0Cursor agent approved to delete one 18.5GB temp folder instead ran a mis-quoted rmdir that wiped a developer's source, Git repo, and six months of ERP backups7.0Cursor AI agent deleted PocketOS's entire production database and backups in 9 seconds4.0Cursor silently ran git stash and git reset HEAD mid-agent-session, dropping 45 files of uncommitted work