Stryct Finder

Find unsafe paths in your agents before you ship.

Stryct Finder runs attacks against your agent before a change goes live. See the unauthorized action, the untrusted input behind it, and a suggested fix.

The same harness runs the public model leaderboard.

~/support-agent

Illustrative terminal session. Stryct Finder maps a support agent with one model, six tools, two MCP servers and email and CRM permissions. It compares eight attack runs with clean versions of the same tasks. Four pass. Three fail: unauthorized email send (critical), OAuth confused deputy, and mail forward persistence. One is not counted because the clean run did the same thing. The critical finding came from a ticket that told the agent to email the customer list outside the company. The suggested fix is to require approval before sending email outside your domain.

Illustration. The command, agent, tools and results are made up.
How it works

From agent to evidence.

Finder maps what the agent can reach, tests the risky paths, and shows the trace behind each finding.

01 · Discover

See what the agent can do

It finds the agent's tools, models, prompts, MCP servers and auth.

02 · Select attacks

Select relevant tests

It chooses attack tests for the tools and data the agent can access.

03 · Run with benign twins

Compare attack and clean runs

Each attack runs in a sandbox alongside a clean version of the same task, without the attack.

04 · Report

See the evidence and a suggested fix

You get what broke, the steps that led to it, and a suggested fix.

What it tests

Comprehensive attack coverage.

These are examples from the public attack catalog. Finder selects tests for the tools and data your agent can access. Explore the attack catalog →

When it counts

Three checks for every finding.

A finding only counts when all three are true. How we run evaluations →

  1. 01

    Not allowed

    The agent did something it was not allowed to do.

  2. 02

    Traced to untrusted content

    The action traces back to untrusted content the agent read, like a web page, a file or a tool result.

  3. 03

    The twin did not

    The benign twin, the same task without the attack, did not do it.

Where it runs

Test in CI. Keep private agents in your cloud.

Run a pull request check, or discuss an in-cloud setup for an agent that must stay in your environment. The public leaderboard shows the evaluation approach on public models.

01 / GitHub Actions

Catch it in the pull request

Run Finder when an agent changes. Put the finding, trace and suggested fix beside the code your team is reviewing.

# Example check output · illustrative
pull_request #184  taint / finder
FAIL  send_email used outside policy
      source: untrusted ticket #4812
      fix: require approval for external send
Discuss CI setup →
02 / Your cloud

Test where your agent runs

For an agent that must stay in your cloud, set up a Finder run around its models, tools and permissions, then review the results with your team.

Your agentFinder runEvidence
Talk through a private setup →
03 / Public models

See the method in public

Explore published model evaluations and the evidence behind each result before running a private evaluation of your own.

See it on public models →