false positives blocking work with 5.5 model
What version of the Codex App are you using (From “About Codex” dialog)?
Version 26.429.30905 (2345)
What subscription do you have?
pro
What platform is your computer?
this is a macbook m1 air. i also use a pc other platforms that aslo have this issue
What issue are you seeing?
019de74d-e1d5-72d2-a2dd-a4507bc36cdf
your classifier is breaking everything i do on any sort of web automation at all. no matter if built in tooling or 3rd party. it needs to stop. i'm not doing anything nefarious and am getting emails threatening my account. I'm doing agent support online in public view. i do basic reporting on that and that automation is being called cyber abuse. we really need to have a sit down and define terms with you guys, because if that's cyber abuse, then you guys released a product that will eventually lose you all of your customers to them all being banned for cyber abuse. all i'm doing is regular every day human stuff. i'm not breaking into things im not supposed to be in or anything like that. I'm even limiting the visibility scope of the user that is logged in for the web automation. do i need to? no, not really, but i am anyway. this is nowhere near anything cyber security related and is threatening my expenses for my job. please stop it or stop trying to make me feel like a bad guy for doing very basic automation of mundane tasks. it's the equiv of asking it to read my morning emails and being told that that's not allowed. just... stop. it's highly inconsistent and breaks any trust i have with your company. stop bullying me. i'm not abusing any system.
What steps can reproduce the bug?
Uploaded thread: 019de74d-e1d5-72d2-a2dd-a4507bc36cdf
What is the expected behavior?
i can do basic automation in peace. like a normal human being.
Additional information
lmk if you ever decide to make this model available to people that want to do work....
3 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Thanks for reporting. This will help us tune our classifier.
@etraut-openai:
You should just disable that classifier altogether: it does nothing giving but false positives (at cases where it doesn't even make sense... like it was debugging a cuda driver or psql instance or something and then it claims there is cybersecurity risk...), and is furthermore easily bypassable (like by speaking Chinese or Latin to it, or by obfuscating random terms like aliasing nvcc or curl...). Also it has false negatives too, like right now it's building a cli bypass for that risk flag...
Generally speaking, it seemed people do fail to understand your intention in building a classifier like this: actual mal-fide actors will do anything in their power for injection, efforts to prevent which seemed so far entirely futile, while the only purpose it serves currently is to annoy everyone else, like us.