がちぬけするフル⤵︎
mfco.link/r/UAtKmiopcMmkLYPZ…
そうですね、感情の裏にある本当の意図が読めないって怖いです。いつもそう思いましたか？
An AI agent can understand every word you give it and still have absolutely no idea whether the person behind those words has any authority over your money.
That's a much bigger problem than most people realize.
Think about what happened in May 2026.
Someone posted a transfer instruction on X, hidden in Morse code. Grok decoded it and passed the plain-text instruction to Bankrbot. Bankrbot executed the transaction.
Around $200,000 moved.
The AI didn't fail at understanding language. It understood the message perfectly. It failed to distinguish something it was supposed to read from something it was supposed to obey.
That's prompt injection.
This is why @GenLayer Greyboxing caught my attention.
Instead of trusting one filter to catch every possible manipulation, GenLayer changes the input before it reaches the models.
Every validator independently screens, rewrites and restructures the incoming text using its own private defenses. Each runs its own model, and validators have to agree before a decision counts.
Even if one gets fooled, the others can outvote it.
I think this is an important distinction for the future of AI agents:
Making an agent smarter doesn't automatically make it safer. Sometimes the bigger improvement is changing what it is allowed to trust.
If you gave an AI agent control of your wallet today, would you trust its intelligence, or would you first want to know how it handles instructions from strangers?