This ☝️ Anthropic's Glasswing project is marketing more than it is a security breakthrough. Giving an AI full source code access and asking it to find bugs is white-box testing; of course it will find something. That is fundamentally different from the black-box reverse engineering attackers actually use to discover and exploit vulnerabilities in the wild.

Yes, the fact that they phrased it as giving $100M in "free" credits (out of the goodness of their hearts, obviously, and not at all to try and hook new customers for their product 😇) did seem a little sus if their motives are truly altruistic. Structuring it that way really smacks of "we've built this so that when we're ready to start charging you for it, all we have to do is flip the switch".

And looks like the level will keep increasing. Next model will be harmful to space and time, and could destroy the universe if we didn't take it seriously.

Well if they start letting AI run the supercolliders, who knows? "Sure! I can generate a black hole for you! Let me get started." 😂

"Oh! You are absolutely right, that black hole destroyed earth instead of the enemy's planet. Let me try one more time."