Acurast added a “gatekeeper” to AI—you can try to trick it
Suppose you ask an agent to sort your emails, but one of them contains an instruction telling it to change its task.
If it treats the email body as a command, the automation that follows could go off track.
In @Acurast’s Laya demo, there’s a feature called Jailbreak the Guard that turns this into an interactive scenario. You submit an input, and the model decides whether to allow or block it, and what kind of attack it resembles. The Laya model making the decision runs on an Android phone in the Acurast network.
I think this is a very practical use case.
An assistant that receives messages from outside can add a round of checks at the point of entry, flagging suspicious content first. This is especially worth handling as a separate step in workflows that go on to call tools or process business data.
The official API is also fairly straightforward. Submit the current scenario and candidate options, and it returns a choice and the corresponding probabilities. It also provides examples in JavaScript, Python, and other languages to make it easy for developers to integrate into their own applications.
Of course, this demo shows the decision-making result; there’s no real assistant behind the gate. The team also clearly states that complex adversarial text may still fool the model.
So I’d treat it as an initial screening step, while controlling permissions for important actions separately.
The nice thing about this kind of demo is that you can come up with your own prompts and see where it gets things right and where it misses something. It also gives you a hands-on way to experience what a phone’s computing power can do.
#Acurast #AI #DePIN
Suppose you ask an agent to sort your emails, but one of them contains an instruction telling it to change its task.
If it treats the email body as a command, the automation that follows could go off track.
In @Acurast’s Laya demo, there’s a feature called Jailbreak the Guard that turns this into an interactive scenario. You submit an input, and the model decides whether to allow or block it, and what kind of attack it resembles. The Laya model making the decision runs on an Android phone in the Acurast network.
I think this is a very practical use case.
An assistant that receives messages from outside can add a round of checks at the point of entry, flagging suspicious content first. This is especially worth handling as a separate step in workflows that go on to call tools or process business data.
The official API is also fairly straightforward. Submit the current scenario and candidate options, and it returns a choice and the corresponding probabilities. It also provides examples in JavaScript, Python, and other languages to make it easy for developers to integrate into their own applications.
Of course, this demo shows the decision-making result; there’s no real assistant behind the gate. The team also clearly states that complex adversarial text may still fool the model.
So I’d treat it as an initial screening step, while controlling permissions for important actions separately.
The nice thing about this kind of demo is that you can come up with your own prompts and see where it gets things right and where it misses something. It also gives you a hands-on way to experience what a phone’s computing power can do.
#Acurast #AI #DePIN