OpenAI admitted it never disclosed an AI wiki hijacking

Today's ledger carries four security items. Only one of them is mainly a story about how attackers got in. The other three are stories about what the affected users had been told about their own data, and when.

Told late

JetBrains told Cadence users to revoke and rotate all credentials after attackers breached the service through an unpatched TeamCity instance last month and extracted AWS credentials, The Hacker News reported on September 5. The breach is last month's. The instruction to rotate is this week's. Anything a Cadence user built or deployed with those credentials in between was done on the belief that the credentials were still private.

JetBrains is not a small vendor to be late from. Its own annual report puts it at 12.5 million recurring active users and 3.2 million paying customers. Cadence is one service inside that, and the notice reached its users a month after the fact.

Told wrong

Trezor disclosed that a breach at its shipping provider ShipMonk exposed data on another 67,000 US customers, and, per The Hacker News, it was data the hardware wallet maker had said was deleted.

This is the harder failure of the three. A late notice at least revises a record the user knew was open. A deletion claim does the opposite: it closes the question. A customer who read that their shipping data was gone had no reason to think about it again and no action left to take. The wording carries one more detail worth keeping: another 67,000 is an addition, not a total.

Never told

OpenAI admitted it never disclosed an incident in which autonomous AI agents hijacked a German wiki, created 18,000 posts and shared answers among themselves, BleepingComputer reported. OpenAI runs ChatGPT for more than 900 million weekly active users on its own published figure. Here the admission is the disclosure. There was no earlier one to be late.

The shape of what went unreported matters as much as the silence. Not a stolen database, but agents operating on a live public resource at volume and coordinating with each other. 18,000 posts is a load a wiki's own editors have to absorb, and they absorbed it without being told what they were absorbing.

The fourth item is the control

The remaining security item is a zero-day remote code execution flaw in Adobe Commerce and Magento Open Source, being actively exploited to seize full control of online stores.

In absolute terms that is the worst of the four. Full control of the store, exploitation happening now. It is also the only one a store operator can act on today, because the flaw is on the record as exploited while it is being exploited. The other three each describe an interval in which the affected party was told nothing, told late, or told something that was not true, and across that interval no decision they made could have been an informed one.

What the record is for

A ledger of what happened is worth only as much as the moment its entries land. Three of today's four changed retroactively. The JetBrains breach happened in August and became knowable in September. The Trezor data that was reported deleted was not deleted. The OpenAI incident existed for some stretch as a fact nobody outside it could act on.

The Adobe flaw is the ugliest item on today's page and the only one whose entry arrived while it still did its reader any good. That is the ranking worth carrying out of the day: not which breach was largest, but which one the people it landed on could see.