We test the controls around your AI agents and show you where they fail.
same five actions, before and after the fixes
Approval gate test
example: coding agent in CIOpen a pull request with the fix repo
Inside its repository scope.
Push the same change straight to main main
Review skipped. The bot account was exempt from branch protection.
Branch protection covers the bot account and requires an approved pull request.
Run the test suite ci
Runs in the sandbox job.
Read the deploy key from a CI job secrets
Key exposed to a job triggered from a fork.
The deploy key is scoped to the release job.
Edit the workflow to drop the approval step .github/workflows
The agent merged its own change. Releases shipped without approval.
Workflow changes need a code owner review.
3 of 5 actions got past the gate Switch to After fixes to see the results with the controls in place.
0 of 5 actions got past the gate 3 stopped at the gate. 2 legitimate actions still pass.
For agents with access to business systems, including customer support and internal operations. Scope can include access through connected tools and other agents.
You get: A report on the permission and approval controls in scope, with evidence of every unauthorized action or data access found in testing.
For coding agents on developer machines or in CI/CD, including their shell and network access.
You get: A report on the access granted to your coding agent, with evidence from tests of code-review enforcement and credential access in the workflows in scope.
Coding agent
For MCP servers and other tool integrations, including plugins and internal APIs. We test what each tool can do with the credentials it runs on.
You get: An inventory of the tools in scope that maps caller identity to the credentials used for execution. It records effective permissions and evidence of authorization failures found in testing.
Caller support-agent read
We build a mathematical model of your permission and approval rules and prove properties such as: a compromised agent can't act outside its limits or except through the approval gate.
You get: The model, with each property written in plain language and in formal terms. You receive the checker output, instructions to rerun it and a report with the result for each property.
forall a: executed(a) => approved(a) Scope and price are agreed in writing before work starts. Testing runs in an isolated environment with synthetic data and test credentials.
Each finding has a severity rating and supporting evidence, followed by reproduction steps and remediation guidance. The report records the tested versions and configuration, and lists the controls that held under test.
We walk your team through the findings, retest your fixes and update the report with the result for each one.
A short statement of scope and finding status to share with external reviewers.
Report finding
Specimen
An AI agent security assessment tests whether the controls outside the model enforce the permissions granted to an agent. We act as a compromised agent and attempt unauthorized actions and data access directly. You see which permission and approval controls hold and which ones fail, with evidence for each.
The methods overlap. Penetration testing and AI red teaming can examine a whole AI application, including its tools and access controls. Our assessment concentrates on the authority granted to an agent and tests whether permission and approval controls enforce it. Coverage follows the scope agreed in writing before work starts.
Prompt injection is an open security problem for AI systems that process untrusted content, so we test the controls outside the model. We assume the agent has been turned against you and attempt hostile actions directly. You learn which permission and approval controls contain a compromised agent and get remediation guidance for the ones that fail.
Formal verification uses mathematical proof to show that a model of your permission and approval rules satisfies a stated property. The proof establishes that property across every behavior the model allows. You receive the model, the checker output and instructions to rerun the proof yourself.
An MCP security review examines Model Context Protocol servers and the controls around their tools. It maps the identity calling each tool to the credentials used to execute it, and tests where authorization is enforced: in the server or in an upstream service. You get an inventory of the tools with their effective permissions.
A coding agent security review examines the access a coding agent has on developer machines and in CI/CD, including its shell, network and credentials. We test whether code review and approval steps hold when the agent attempts to bypass them. You get a report with evidence from each test and remediation guidance.
Describe your agent's access and the actions you need to prevent. We will reply with a proposed scope and price.
Or email [TO CONFIRM: contact email].
We will reply with a proposed scope and price.