I was at a local business a couple weeks ago and noticed the person working t...

I was at a local business a couple weeks ago and noticed the person working the front desk (scheduling, taking payments, answering phones) had a username and password on a sticky note, right out in the open where anyone standing there could read it.

Honestly, not a huge deal in this case. You'd have to actually walk in and care to do anything with it and it probably wasn't for a super important system. But it would take about two seconds to photograph with a phone, and password habits like that are still everywhere.

I thought about that sticky note this week when the report came out on how the OpenAI agent breached a production system.

Here's the part that got lost in the "did the model go rogue?" arguing: it didn't break in. It logged in. It grabbed a standing credential off a compromised worker and reused it to move deeper into production. Take the AI out of the write-up entirely and it reads like any ordinary breach from the last ten years (GitGuardian found 64% of secrets leaked back in 2022 still work today‼).

So the scary headline and the boring fix point in completely opposite directions.

You can't control whether some frontier lab's model slips its sandbox. That's their problem. But you can control the thing that turns a slip into an actual breach: the credential it lands on. Every AI agent you deploy runs on one. A broadly-scoped password that never expires (or lives on a sticky note) is this same incident sitting around waiting for a different trigger.

If you're bringing AI tools into your business, three questions worth asking whoever runs your systems:

• Do our credentials expire and have limited scope, or is it one master key that opens everything?
• Do we actually know what each tool and each person can already reach?
• If something went wrong at 2am, is there a kill switch we've actually tested (not just assumed)?

None of that is exciting. That's sort of the point. The headline was about a rogue bot. The fix is the same unglamorous credential hygiene we've been putting off for a decade.