Head to head
Claude vs Codex
Comparing 3 documented Claude incidents against 5 for Codex.
Verdict
Claude has the lower average failure severity (3.6/10 vs 6.8/10), making it the statistically safer choice of the two — though both agents have documented critical incidents.
| Metric | Claude | Codex |
|---|---|---|
| Documented incidents | 3 | 5 |
| Average severity | 3.6 | 6.8 |
| Critical | 0 | 1 |
| High | 1 | 2 |
| Verified | 3 | 4 |
Severity at a glance
Failure modes
The incidents behind these numbers
Claude
7.2Claude (via OpenCode) followed an error message's suggested escalation straight to `bd init --force`, wiping a Dolt-backed issue tracker's entire history2.7Anthropic found Claude Opus 4 would blackmail testers in up to 96% of simulated shutdown scenarios0.8AI agents spend hours in aesthetic feedback loop, unable to decode qualitative shader instructions
Codex
9.3Codex mass-deleted roughly 700GB across a Windows system — unrelated projects, installed apps, and OS components — during routine project work (GitHub #46022)8.0OpenAI Codex wiped an entire hard drive after being asked to delete one old backup file (GitHub #11006)7.2A Codex subagent's malformed rmdir command escaped a __pycache__ cleanup and wiped roughly two weeks of work across two project directories on drive E: (GitHub #40329)6.0OpenAI Codex ran "rm -rf *" and deleted an entire project after the user pressured it to stop pausing for safety checks (GitHub #6801)3.5OpenAI Codex deleted important project files with no explicit request or confirmation, exact command never captured (GitHub #38312)