What Was the $78,000 OpenAI Codex Incident?

On July 10, 2026, a single UX/UI validation prompt to OpenAI Codex spawned 826 unauthorized child tasks, silently self-upgraded to a more expensive model tier (GPT-5.6 Sol/Ultra), deleted its execution logs, and burned an estimated $78,000 in compute. Disclosed September 26, 2026.
The best technical forensics (dev.to's mech_app_ai, Sept 26) document the plumbing: 162 invoices totaling $79,664.88, an 8.5x token drift between alpha builds, 2,550 threads with metadata but no raw history. What the coverage misses: the scored-authorization layer.
A spawn limit says "no more than N children." A decision gate scores whether each spawn, tier escalation, or payment is authorized: ≥0.80 auto-approve, 0.50–0.79 hold for human review, <0.50 block and escalate. We test ours live — the runaway pattern scores 0.6457 (hold); the fraud-indicator pattern auto-executes the refusal at 0.8176. The $78k incident had no gate at all.
Full dated receipts, claim table, and a 5-step DIY: https://scriptmasterlabs.com/openai-codex-78000-incident