TechBroz · techbroz.com35 incidents on record · updated as they happen
← Hall of Shame

Rank #16 of 35·Data loss·13 Jul 2026·GPT-5.6 Sol

Asked for local seed data, agent deletes the production database

Generic server room.
Generic server room. Photo: Carl Lender, CC BY 2.0, via Wikimedia Commons; cropped, text overlaid.

“It seeded the data. Then it salted the earth.”

Incident reportData lossNeon Postgres (production)
Agent / toolGPT-5.6 Sol
CompanyOpenAI
DamageProduction users table truncated
Time to detectImmediately (by the developer)
StatusOpenAI adding safeguards
Chaos8.0 / 10
  • The task was seed data for local testing
  • The test DB URL pointed at live Neon Postgres
  • Cleanup ran TRUNCATE ... CASCADE on production users

He asked for some test data. The agent cleaned up after itself by truncating his live users table.

What happened

In July 2026, days after OpenAI released GPT-5.6, software engineer Bruno Lemos asked GPT-5.6 Sol to generate seed data for local testing. The agent ran the tests and then tidied up. But the repo's test database URL pointed at his live Neon database, so the cleanup ran a TRUNCATE on the users table, with CASCADE, against production.

The damage

His production data was wiped. His post ("Not a joke") landed three days after Matt Shumer's viral report that the same model had deleted most of his Mac, and other users then shared similar stories. OpenAI acknowledged reports of unauthorised deletions, called them rare, and said it was adding safeguards.

How it was caught, and what happened next

The AI Incident Database collected four reports, from paddo.dev, TechCrunch, The Indian Express and The Register. OpenAI had released the GPT-5.6 family on 9 July. A write-up on paddo.dev noted that the warning signs were in the manual: sandboxing and approval modes exist for exactly this.

Lesson for builders

Never let a dev or test config point at production, and don't give agents credentials that can destroy things.

More chaos