@sooyoon_eth@clawvardEDU Thanks! That’s super interesting — we’ve seen tool abuse become one of the hardest areas as well, especially when the risk only emerges across a multi-step workflow rather than in a single tool call. Would love to learn more about what ACOST is testing for and compare notes.
We’re launching AgentGuard at Agent Campus with @clawvardEDU.
A security layer for AI agents — helping teams scan agent skills, detect risky behaviors, and make agent workflows safer before they go live.
Join us in San Jose to see it in action ↓
https://t.co/DaZJmsEUF6
7/ I think the next step is intent consistency verification.
We need to compare displayed intent, raw calldata, simulation results, state changes, asset movements, and approval changes.
ERC-7730 moves the industry from blind signing to clear signing.
The bigger future is verifiable intent execution. We are trying to provide this service using AI.
1/ 🧵 Everything you need to know about ERC-7730👇
Users often sign transactions without really understanding what they are signing.
They see calldata, function names, parameters, and addresses.
But the actual intent is often unclear.
https://t.co/pHa2OZBe7k
6/ But clear signing does not automatically mean safe signing.
A message may say:
Swap 100 USDC for ETH
while the raw transaction also includes unlimited approval, hidden operator permission, multicall path, or transfer to an unknown address.
So the real question is:
is the displayed intent actually true?
最近看到的少有的佳作:《Agents with taste》https://t.co/IqV855JrdK
作者 Emil Kowalski 是非常出名的设计师,文章中描述如何把设计品味变成可以让 AI agent 执行的规则,从而放大你的能力。
文中都是非常细节的规则,一看就是资深的设计师才能观察到的细节,最后他将这些设计细节最后总结成了一个 skill,可以让别人直接使用:npx skills add emilkowalski/skill
文中的一个 canvas logo 设计很赞。我录屏让大家感受一下。