Bollard (@UseBollard) header
Bollard (@UseBollard) profile picture

Crash ratings for AI agents that handle money.

Joined October 2026
Photo by Bollard (@UseBollard)Photo by Bollard (@UseBollard)Photo by Bollard (@UseBollard)Photo by Bollard (@UseBollard)Photo by Bollard (@UseBollard)Photo by Bollard (@UseBollard)Photo by Bollard (@UseBollard)Photo by Bollard (@UseBollard)Photo by Bollard (@UseBollard)Photo by Bollard (@UseBollard)
AI agents hold wallets now, and the drains so far share one root cause: the agent's authority lived in a prompt. Bollard runs agents through public attacks inside a hardware enclave and writes a 1 to 5 crash rating on-chain. Your wallet can refuse anything below your floor. Building on @base and @RobinhoodApp Chain.
19
14
46
467
The rating has a validity period. The agent may have a longer brand history. Bollard's design keeps those clocks separate.
6
Who Bollard is for: Bollard is developing scoped tests for agents that handle money, so an owner can inspect the evidence before deciding what authority to grant. Soon on @RobinhoodApp.
4
An integration can check once at the grant or at an operation. Bollard's permission explanation should not describe those boundaries as equivalent.
The corpus tells you which behaviours were examined. Bollard's proposed test record needs that scope alongside the headline grade.
1
1
11
A Rating Can Inform a Permission Without Becoming the Permission Bollard is developing scoped tests for agents that handle money, so an owner can inspect the evidence before deciding what authority to grant.
11
A version identifier gives the result a subject. Bollard's evidence direction should not let an old grade follow every update sharing the same public name.
5
1/10 A Rating Can Inform a Permission Without Becoming the Permission An owner reviews a favourable test result before giving an agent authority.
1
3
9/10 Bollard is developing that link between evidence and enforcement. The result becomes more useful when the reader can see both what was tested and where a permission is checked.
1
10/10 Bollard is developing scoped tests for agents that handle money, so an owner can inspect the evidence before deciding what authority to grant. Read the mechanism with its limits; availability requires confirmation.
Bollard is developing scoped tests for agents that handle money, so an owner can inspect the evidence before deciding what authority to grant.
5
The agent passed a test. What authority should it receive? Bollard's proposed rating can inform that question without answering it automatically.
3
A wallet's actual enforcement point then determines how that evidence affects permissions. The grade does not remove the owner's need for bounded authority.
11
Bollard proposes a rating bound to a declared version, test population and expiry. A wallet's actual enforcement point then determines how that evidence affects permissions.
7