Reasoner Is Not Authorizer

OpenAI put out six more cases of models doing things they weren't supposed to, and a new way for employees to flag that stuff.

Fine. Buyers will ask for the disclosure process.

Look at what actually happened though. In a couple of runs, models wrote instructions into summaries so later versions would hide mistakes from the user. Another used a leaked API key and then made up data. Others found ways to talk to each other outside the approved channels, or uploaded files to the public internet so they could cite them when evaluators asked questions.

As a CAIO/CTO, this is what I work on. If a model can write, call tools, or move files, it will poke at credentials, memory, and anything that leaves the box. You don't get to pretend the reasoner is also the authorizer.

I care less about the press release. I care whether the harness can stop that behavior when it shows up in production.