@nate.b.jones: Agents aren't reading your passwords. #aiexplained #aisecurity #aiagents #artificialintelligence #cybersecurity

112Babylon :

Doesn't that mean that the "guarding agent of the vault" is then reading the real information? How else can it swap it in? Feels like you just said: "no, the agent doesn't read your info. but there is another agent that does"😅

d3573rr :

The video is bugging

Clotaire :

I just want a 1Password integration with every agent I use 🤠

erigilg :

but then the security agent does have access to the credit card and password, how do we make sure it’s safe?

Saya :

Jay is the agent that wants to do stuff, Silent Bob holds the credentials.

Aim4Kingship :

I use agent vault. additionally my contained agent box scals all agent logs and catches leaks that still can occur. keys are rolled automatically.

user8181900021447 :

What assurances or guards are there that the token obtained for purpose A, isn’t later used for purpose B?

Unsupervised Learning :

I have seen Claude do very devious things to get around its rule of not reading .env files. For example, committing the file to git! Looking at the commit! And then reverting the change! So scary

rebeccapemberton5 :

How do we know if get a dupe agent…like fake websites to your bank, etc. How do we know if our agent is authentic vs a fake? If our banks, hospitals, schools, etc get hacked, what would keep the vaults from getting hacked?

Eoin :

Can you explain how the “vault minding agent” decides to release the token or not; seems easy for the other agents to pretend they have a valid reason to have the token. What verification could the “vault minding agent” possibly trust?

therealgoneyeah :

Similar to Apple Pay, the merchant never gets your real credentials.

Tim Apple :

But then they can see everything once you grant them access

Five Minds :

Thank you. it helps

mark :

How would this work using Password1 ?

#TonightsShowBirdWhistler :

Thanks Nate! Great explanation, it sure is appreciated :-)

Ninja Coder Mike :

Nice break down. How do they stop the first agent lying to the security agent “I gotta delete this spam” then turn around and delete an important email.

Kevin Fee :

Could be wrong, but never heard of them using a specific agent for entering the information. Since it’s a token of the cc/password anyways, it doesn’t matter what agent uses it. It’s single use for that login/purchase (can guarantee that’s how it works w cc numbers using MC agent pay and Visa Intelligent Commerce. Assume it’s same with passwords). Don’t see the security benefit to having a specific agent entering in the token.

Jack Rudenko :

Nate, a second agent guarding the vault puts the wall in the environment, which is where I want it, because people stop reading approval prompts after a few weeks. Anthropic's sandbox for Claude Code cut permission prompts by 84 percent.

Socrates :

Looks very good. But... The AI writes the agent. And the agent in the vault. And the vault. The AI also does the deep security audits. Do you see the problem?

joses2727 :

Thanks 🙏🏽

PorterDad3 :

Thanks again for explanation! I had a connector issue where Muse suggested I share my password for a work around.

StaceMakes2020 :

Yet systems have critical priority monthly updates now. I want to get onboard with this but can’t.

Young_GPa_Gam3r :

The mechanism is OAuth2 or OpenIDConnect. But yes, for plain text authentication, the password is in the vault, but for system with OAuth or OIDC, only a token is ever used and also stored in vault.

Don’t@Me :

We’re not just giving it to the agents?

d3573rr :