Further secure access-token-feature for use with AI agents
The biggest security risk when using the new token access feature for ai agents is, that the agent tries to exfiltrate passwords because of prompt-injection attacks. This in general is hard to prevent. But maybe there is a way to add "honey-pod-passwords" for which there would be no light reason to be accessed. If they are accessed by an agent gone wild, the pass - cli immediately locks and warns the user.
1
vote
Luke
shared this idea
-
T
commented
This complementary request moves the security boundary earlier: instead of only detecting suspicious access after an agent can retrieve plaintext, Proton Pass would broker the login so the agent never receives the credential. Canary detection and brokered autofill could work together.