Colby Farley

These articles and tweets are my own opinion and do not reflect the view of my employer.

I am a security enthusiast, researcher, and multi cloud fixer

[Colby Farley @Taco_Rocket]

Azure persistence does not always look like a backdoor.

Sometimes it is a Function App, WebJob, ML job, Container Apps Job, or VM Extension that can run again later.

HarrierOps Azure v1.2.0 helps pull that forward.

Can you spot the re-entry path?

[Colby Farley @Taco_Rocket]

@neondatabase Why not just throw provider support for Terraform or OpenTofu? The CLI clearly follows their naming convention. Would also be another dependency to not install

[Colby Farley @Taco_Rocket]

@HackingDave This right here! Hyped it up. It's a decent upgrade but it ain't the end of the world.

[Colby Farley @Taco_Rocket]

@jxnlco @jsnnsa @TAbrodi No he's using Gavi and ravi /s

[Cheng Lou @_chenglou]

Few people know the secret, which is that Codex actually has a giant unfixed bug that sporadically resets usage limits. Every time this happens Tibo announces that he did the reset

[Colby Farley @Taco_Rocket]

@HackingDave The hold its ground mechanism is kinda silly. You can just have it delegate those items to Haiku and it will do it still

[Colby Farley @Taco_Rocket]

@peter_szilagyi Just tell it that if Claude doesn't, Codex will. It usually just runs it. The feedback mechanism just wants you to provide evidence

[Colby Farley @Taco_Rocket]

@thsottiaux Now can we get better subagent management? The back off retry, slows workflows.

[Colby Farley @Taco_Rocket]

@gdb Hi I reported an issue and tried to submit a PR to expose an endpoint to help clean up the subagent pool. Happy to discuss but if we could start there it would help me a lot!

[Colby Farley @Taco_Rocket]

@jxnlco @encrypted lol no way. Have you seem the amount of stuff Claude flags for cyber

[Colby Farley @Taco_Rocket]

@jxnlco The one I just raised an issue for. Happy to submit a PR!
https://t.co/pJ1P8Un936

[Colby Farley @Taco_Rocket]

@thsottiaux can we get a bit more automated/proper subagent handling in the GUI or heck just expose /agents as an api to have a script codex fires off to kill them? Happy to use the CLI if this feature is here but this is the issue I'm running into

[Colby Farley @Taco_Rocket]

@labibrahman @steipete Dude the fact compaction doesn’t take 5 mins I’m here for it

[Colby Farley @Taco_Rocket]

@Dimillian Is there any way to have codex threads bleed into chat app or at least have a “send to ChatGPT” I’m not even trying to get around codex limits mind you, but having an idea passed to ChatGPT like an outside “is this dumb” opinion is great. I just don’t want to have to type again

[Vlad Sazonau @vladsazonau]

am I safe now?😅

[Colby Farley @Taco_Rocket]

@nicdunz My favorite

[Colby Farley @Taco_Rocket]

@nicdunz Because I’ve used Claude as a code reviewer and I’m gonna be real. It’s more skeptical if you tell it Claude did it (any llm actually) but if Claude signs the task as codex. Codex don’t care

[Colby Farley @Taco_Rocket]

@jxnlco Can I just get the ability to clean code review from codex? It’s too much in the weeds. I just want a chapgpt mode essentially where it takes off the implementer hat for a moment

[Colby Farley @Taco_Rocket]

@BenjaminBadejo @steipete @thsottiaux @OpenAIDevs @OpenAI That’s a lot of words for saying that you basically don’t want to pay for API costs when it comes to your “consultations” or “solutions”

There’s also a lot of tools that help calculate API costs to usage. You can use that and get a rolling estimate. Then extrapolate from there

[Colby Farley @Taco_Rocket]

@BenjaminBadejo @steipete @thsottiaux @OpenAIDevs @OpenAI So basically you just want more subsidized compute? You do realize at the $200 plan you’re already getting way more than what the API usage charges? Which is wild because I have to do high and fast to get close to below half. I’d look for efficiencies first in what you’re doing

[Colby Farley @Taco_Rocket]

@bcherny @sickdotdev It’s the chat app. It’s kinda wild you can burn through your usage with it. Even if it was counted like half, I’d use Claude more. Made a code review plan with claude and it ate 79% of my 5 hours. This was with chat and only generated the doc once