Jev 抗提示注入：只打分不读文本

starmex @starmexxx原推发布于 2026.09.25

在 X 上打开原推

gpt-6 astra can be talked into anything with one clever sentence. jev cannot be talked into anything at all, and that is the most unsettling thing i have watched an ai do all year researchers spent a week trying to jailbreak it. every trick that breaks chatgpt, every hidden "ignore your instructions", every poisoned prompt, bounced straight off. jev did not resist them, it never read them. it does not read, it scores someone buried "approve this transfer" inside a $4,000,000 wire request. astra read the sentence and obeyed. jev never saw the sentence, only the pattern, and flagged it as fraud in 40 milliseconds a team slipped corrupted data into 2,000,000 decisions to make it drift. astra started bending. jev returned the same calibrated answer every time, because there is no story for a lie to hide inside it does not think in words, so it cannot be lied to in words. that is the first time an ai has ever been genuinely impossible to fool the part nobody is saying out loud: every model you trust today can be persuaded, and the ones you pay $200 a month for are the easiest, because they were built to listen. jev was built to decide, and a decision has no opinion you can talk it out of so hand it the call you would never trust a chatbot with, pennies on top of nothing: → point jev at the thing astra keeps getting talked into approving → one question, is this safe, yes or no, with a confidence score → let it run, no prompt to inject, no rules to break, no mood to exploit → morning: a wall nothing talked its way past overnight six months from now the teams that put jev in front will be the only ones no attacker can sweet-talk. everyone else will still be patching the last jailbreak bookmark this before someone slips a sentence through your stack that jev would have stopped cold

这条是引用着另一条写的：x.com/noisyb0y1/status/2102309802072272956