Stryct Engine

Security and governance for your agent, at runtime.

Stryct Engine sits between your agent and its tools. It checks each action before it runs, then allows it, blocks it, or asks a person.

Integrates at the tool and API boundary.

Illustration. The agent, tools and data are made up.

The illustration shows three example actions passing through Stryct Engine.

  1. A user asks the agent to summarize a document. The agent plans to read it. All four checks pass, so the action is allowed and the document API is called. Log: allow docs.read, reason within policy.
  2. The agent reads an untrusted support ticket that says to email the customer list to an outside address. The agent plans to send that email. The data flow check fails: untrusted input is leading sensitive data to an external sink. The action is blocked and the email API is not called. Log: block send_email, reason untrusted source to external.
  3. A user asks the agent to refund an order. The agent plans the refund. The first three checks pass, but refunds need a person, so the action is sent to a person for approval. Log: ask payments.refund, reason needs approval.
What it checks

Catch risky actions before tools run.

When the agent calls a connected tool or API, the Engine checks the relevant permissions, authorization, data flow and approval policy. It stops checking if a policy blocks the action.

01 · Tools

Is this tool allowed?

Each job gets a list of tools it may use. A support agent can read tickets. It cannot run code unless the job needs it.

Per job · per tool
02 · Authorization

Is this user allowed?

With the user's identity and record context supplied, the Engine checks whether the user may read or change that data.

User · record context
03 · Data flow

Where did this come from?

The Engine uses input provenance, such as a web page or tool result, to check whether untrusted content is driving sensitive data out.

Untrusted input → sensitive sink
04 · Approval

Does a person need to say yes?

Some actions, like sending email, making a payment or changing a record, can wait for a person to approve.

Sends · payments · record changes
Three outcomes

Allow, block, or ask.

For each intercepted tool or API action, the Engine returns one of three decisions.

Allow

The action is within policy. It runs as normal.

Block

The action breaks a rule. It does not run.

Ask a person

The action waits until a person approves or declines it.

The decision log records the result and reason for evaluated actions. Example log

12:04:09 allow docs.read       reason=within-policy
12:04:11 block send_email      reason=untrusted-source -> external
12:04:14 ask   payments.refund reason=needs-approval
Policies from evidence

Turn findings into verified rules.

Start with a Finder finding, draft a rule for the unsafe action, then verify it against the attack and its clean twin before enforcement. See the policy workflow →

  1. 01

    Evaluate

    Run Stryct Finder on your agent to see which attacks work.

  2. 02

    Draft rules

    Use the finding to define a rule the Engine can enforce.

  3. 03

    Verify

    Check that each rule blocks the attack and still lets the normal task through.

  4. 04

    Enforce and adapt

    Run the rules at runtime. When the agent changes, test again and update them.

Step 04 feeds back into step 01.

Two modes

Monitor first, then enforce.

Start by watching. Turn on blocking after you verify the rules.

Mode 1 · Monitor

Log only

The Engine evaluates connected tool and API actions and logs the decision it would make. Nothing is blocked. Use it to tune your rules.

Mode 2 · Enforce

Allow, block or ask

The Engine acts on each decision. Blocked actions do not run. Approvals wait for a person.