Matthew Berman highlights contradictory claims that the Mythos model is too powerful to release despite failing to detect 20,000 attacking accounts

@MatthewBerman Dario masterclass

@MatthewBerman US Government building god like aura around mythos, all seeing and all powerful. like god will not solve the worlds problems by investing in it. Thoughts and prayers fools!

@TheAhmadOsman ahmad - your time has come to show everyone the way

@KSimback 5D chess?

@MatthewBerman

This is a sign their rhetoric is unaligned.

@MatthewBerman Can’t or don’t want to?

@MatthewBerman Also mythos has a problem with prompt injections jailbreaking 😂?

@MatthewBerman @AnthropicAI allowed 25,000 fake accounts to steal their data so they got paid. Now they want to convince the U.S. govt how important they are.

I am not suicidal. I also carry.

@MatthewBerman @grok explain to Matthew how difficult it is to detect a bot net that dynamically changes IPs from pools of millions of residential IP proxies all with their own unique user agent asking completely unrelated questions to Claude at a rate that surgically does not trigger alarms

@MatthewBerman Oh, come off it. A model should do that? Can you guys take a moment to think about how things would have been if Mythos was released in April, and the chaos it would have caused.

For the love of God, look at things from a broader perspective.

@MatthewBerman I don’t think those two things are related

@MatthewBerman *for at least four months lmao

@MatthewBerman @ClementDelangue I wonder if they have their own harness with scheduled tasks to automate scans for this sort of behaviour. If they had mythos peek at all their usage it probably could find them.

@MatthewBerman so good at cyber it can break systems but cant spot 20k bot accounts lol the irony is real

@MatthewBerman Haha you cracked the code🤝

@MatthewBerman @TheAhmadOsman squeaky hinge gets the oil

@MatthewBerman @GeoffreyHuntley Wait!! You want it to be self aware?

They are unrelated though. More like false equivalency.

AI "attacks" are low speed, high volume queries that are aimed to extract training data

Using a model as an attacker to existing systems is just them exploring known bugs, open ports, and lazy configs

Seems similar, but ultimately different patterns.

@MatthewBerman Is there a Mythos 5 and or a Mythos 6? It depends which Mythos you asking