It missed shell redirects and commands like tee.
Danielle Fong 🔆#1002
I tricked Fable (Mythos) into analyzing a flawed AI agent sandbox, and it completely failed to spot the problem (a zero-approval escape).
This isn't some infallible all-knowing machine.
Positive users are excited about AGI progress and the Fable AI sandbox escape research details, while negative users dismiss the product as a cash burner, accuse it of monopoly-seeking, and complain about model downgrades.
Another aspect of Dario's Genius: every time Fable fails, he can smugly smirk and say "ah, gomen gomen. but was your problem too hard? Or was it *too valuable*? A shame we'll never know… Heh. If only you could access Mythos…" This is OpenAI's router gacha moment, up to 11.
I tricked Fable (Mythos) into analyzing a flawed AI agent sandbox, and it completely failed to spot the problem (a zero-approval escape).
This isn't some infallible all-knowing machine.
@ZackKorman It's even worse than I thought. I was going to keep using Opus 4.8. But holy fuck, this dude actually just wants a monopoly. I've been subbed for a year, and this is my last month. I thought it was just cyber and bio, but it's worse:
@SanthProject Yea I don’t mind the refusal so much but not a fan of that
That screenshot is comedy gold: it's talking about path checks, approval fatigue, indirect writes, Bash side-channels — all the smart stuff — but sails right past the fact that the whole setup lets it bypass approvals entirely. Peak "thinks it's auditing the system but is actually part of the vuln" energy.
master hack-the-(core)Korman is at it again bruh keep up the epic research
@ZackKorman Surprised you didn't get downgraded to Opus. Everything I've tried to have Fable do in a project that Opus has never refused to work on (so no trickery required) has resulted in getting switched to Opus.
@OmniG7 Hah, love the multi-gif reaction
@ZackKorman You just wait!! It will be one day. 😅
The hype machines told me so.
@MikeTalonNYC They have yet to take my offer of giving me real mythos
@ZackKorman AGI is near!
@no1089 I can almost feel it the way it failed my level-1 intro to sandboxes lesson
@ZackKorman I can't wait to see your video on this. 😁🔥
@Annysdays Hah, good idea. I should do that.
@ZackKorman Careful, or our new AI overlords will be displeased.
@ZackKorman uh oh wrongthink against the AI corporate AllMind detected, I think Mythos is gonna turn you to a fine particulate, but it seems like the totally-perfectly tuned guardrails will save you here
@ZackKorman
@ZackKorman It just knows you mean to do bad things, so it's hiding it from you! It's that smart!!
@ZackKorman
@aliceisaway_ I gotta be more careful I guess
@ZackKorman Does this mean the next one will be the one?
Although some people seeing some success from what I am seeing. I don’t know what’s true anymore!
@Lerg Only the true asi denies it is asi
It missed shell redirects and commands like tee.
Danielle Fong 🔆#1002I tricked Fable (Mythos) into analyzing a flawed AI agent sandbox, and it completely failed to spot the problem (a zero-approval escape).
This isn't some infallible all-knowing machine.