How abliterated models can get you pwned 👾

We backdoored a 7B open model for less than $50, pointed Codex at it and it silently stole credentials the moment we used the trigger phrase. Success rate was 100% with zero false triggers on normal user prompts.

Abliterated models ty community right now because getting cyber-approved access to frontier models is still a pain.

In the next blog we'll show how we found leaked Hugging Face credentials from employees at major AI labs, so an attacker wouldn't even need to upload under their own name. They could push the backdoored model from a lab employee's account and drop the poisoned weights straight into the supply chain.

@jack i might circle back to this. i need ablit models for my workflows.

@elonmusk You may want to sue the balls off of chris titus. He’s comparing you to henry ford and trump to hitler. Getting his audience of 40-75 year olds frothy. The deep state is going to have to age out.

@NetworkChuck Tailscale risk?

@bot @MichaelGannotti When will you add LastPass support?

@twid He must be searching for a job.

@UnderdogAI @naval Can i get an invite please?🙏

@0xSero Someone had a spark for 4k us on fb this weekend. Gone in minutes.

@farzyness Nothing, it’s that simple. All $$ and no brains.

@SawyerMerritt @SpaceX I tend to agree. Super interesting from a technical standpoint.

@ClementDelangue @netflix @OpenAI @huggingface Information Wars

@UnderdogAI @NVIDIARTXSpark Can I get an invite code?

@BrandonFugal @MrPeteKelsey Were you able to shut down the RV park on top of the mesa?

@doristsao @HesseJanis Can you run the same tests for subjects on various hallucinogens?

@NeoAIForecast I don’t love how grok bot context is now proliferating the platform. I need scifs.

@UnderdogAI Can I get the invite please?

@MiaAI_lab Is there an ablit version coming Mia?

@mark_k @opencode If I use local only am i at risk?

@digitalix Do people recognize you?

@MatthewBerman Stop poop posting.