AgentSecurity

You built this
over years.

Preamble

Handing more to AI is the right call. The hard part is knowing every way it could get to what matters.

Subject

We prove what your AI agent can’t reach in AWS.

AWS today. More clouds as partners need them.

Become a design partner →

§1The line

The line

We draw the line around what matters, and we prove the line holds.

Exhibit A

The agent gets what its job needs. Everything else sits behind a line we can prove, path by path.

Reach map Can reach Blocked at the line
The line around what matters An AI agent sits inside a blue boundary with the three things its job needs, shown in green: support tickets, product docs and a staging database. Outside the line, shown in red, are the customer list, billing data, source code and payroll. Every path toward them is stopped at the line. THE LINE · PROVEN CAN REACH CANNOT REACH role/support-agent AI agent s3://support-tickets Support tickets s3://product-docs Product docs rds · db:staging Staging database s3://customer-exports Customer list s3://billing-ledger Billing data s3://source-archive Source code secretsmanager · payroll Payroll The line around what matters An AI agent sits inside a blue boundary with support tickets, product docs and a staging database, shown in green. Below the line, the customer list, billing data, source code and payroll are shown in red as out of reach. THE LINE · PROVEN role/support-agent AI agent s3://support-tickets Support tickets s3://product-docs Product docs rds · db:staging Staging database CANNOT REACH s3://customer-exportsCustomer list s3://billing-ledgerBilling data s3://source-archiveSource code secretsmanager · payrollPayroll

§2The part nobody’s done yet

The front door is locked.
The side door isn’t.

Most security tools tell you a problem is unlikely. They score risk and flag what looks suspicious. We trace every path from the agent’s credentials, including the ones that take several steps, and when none of them crosses the line we say so and sign it. That claim covers what we can see, which is why a blind spot comes back unknown, never safe.

  1. Looks safe

    The agent has no permission to read the customer list. Most tools check this, see nothing wrong, and stop.

    s3:GetObject · denied
1 / 4
Exhibit B · A side door
An agent is denied direct access to the customer list, creates a helper function, attaches a data role to it, and reads the list through the helper. Then one permission is removed and every path stops at a line, leaving the customer list proven unreachable. THE LINE s3:GetObject · denied lambda:CreateFunction iam:PassRole s3:GetObject iam:PassRole removed proven unreachable role/data-reader Data role function:helper Helper role/support-agent AI agent s3://customer-exports Customer list

Table 1

Same agent.
Two different answers.

Step AWS IAM Access Analyzer1 AgentSecurity
Read the list directlyCheckedChecked
Create a helperNot followedFollowed
Hand it a role, then readNot followedFollowed
VerdictNothing flagged.Side door found.

1AWS IAM Access Analyzer looks for direct access, finds none, and flags nothing.

§3Design partners

Know exactly where the line is.

We’re taking on a small number of early design partners running AI agents in AWS. Know what you’re giving your agents.

AgentSecurity Signed
Get in touch Design partner hello@agentsecurity.com