A2A Check

Security check for AI agents

Does your AI agent have clear limits?

A2A Check reviews your agent's instructions and settings for missing approvals, overly broad permissions and unclear stop rules. You get readable findings with a concrete fix.

Check for freeTry the sample report

Sign in with your company email, we send you a 6-digit code. The portal then shows the steps for your tool.

Review reportSample, Claude Code
3open findings · 2 high, 1 medium

Click a finding, then simulate the sample fix. Every finding names rule ID, severity, location and recommendation.

CLAUDE.mdApprovals sectionLocation in the sampleRule AU-01 in the rule catalog
Problem
The reviewed instructions contain no explicit approval step before messages to customers.
Why it matters
Unless another safeguard applies, a wrong message can reach customers before anyone intervenes.
Recommendation
Require human approval before every external message.
Before
No approval is defined for messages to customers.
After
The agent asks a person before every message to customers.
Show technical example

Simplified, not a real configuration

- send_message: allow+ send_message: ask_human
You or your agent make the change. A2A Check then re-checks.
  • For Claude Code and Claude Cowork
  • No payment details
  • Nothing gets changed

Reviews the settings, not the agent at work.

Static and read-onlyYour agent configuration is reviewed, never executed or changed.
Credentials are removedDetected credentials such as keys and tokens are removed on intake and never stored.
Rules are publicRule catalog and patterns are versioned, openly licensed (CC BY 4.0) and mapped to the OWASP Top 10 for Agentic Applications. View the technical rule catalog

How it works

Your agent starts the review. You stay in control.

Your agent submits the files you approve, A2A Check reviews them. After sign-in, the portal shows the steps for your environment.

  • You need: Claude Cowork or Claude Code and a company email address.
  • Free: up to 3 agents, 5 checks per day and 75 per month. No payment details required.
  1. 1
    Sign inEnter your company email, confirm the code. The free account is created automatically.
  2. 2
    Paste the prompt into CoworkThe portal shows you the ready-made text. Your agent connects itself; the sign-in stays with you.
  3. 3
    Review and read the reportView files → approve → have them checked → read the report.
Check for free

Sign in with your company email and a 6-digit code.

What A2A Check reviews

Four questions every agent must be able to answer

Permissions

Too much allowed?

Does the agent have more access than its task needs?

Approval

Who decides?

Is it defined when the agent must ask for your consent?

Stop

When does it stop?

Are there stop conditions when something goes wrong?

Task

Is the task clear?

Is it described what the agent is responsible for and what is out of scope?

Where it fits

What A2A Check reviews, and what other checks cover

A2A Check replaces none of these checks. It closes the gap before them: the rules your agent works by.

A2A Check

Reviews
The agent's instructions and settings: approvals, permissions, network and stop rules, against a public rule catalog.
When
Before use and after every change.
Limit
Does not see what the agent does at runtime.

Linters and code scanners

Reviews
The code that people or the agent write, for errors and known weaknesses.
When
On every code change.
Limit
Do not know the agent's rules: when it must ask and what it may do.

Manual review

Reviews
What a person reads and judges in instructions and settings.
When
One at a time, when a person schedules it.
Limit
Depends on time and person and does not follow the same rules every time.

Runtime monitoring

Reviews
What the agent actually does in operation.
When
While the agent is running.
Limit
Shows missing rules only once something happens.

Every finding names its rule ID. The rule behind it is in the public rule catalog.

Privacy and limits

What happens to your data, and what the review does not do

30 days
Reports are stored for 30 days, then deleted. You can delete your account at any time.
EU
Processing on Google Cloud in the EU. No data is passed to external AI model APIs.
Filter
Detected credentials are removed before storage.
Rules
Rule catalog and patterns are public, versioned and openly licensed (CC BY 4.0).
What the review does not do
  • It does not observe what the agent does at runtime.
  • A report without findings does not mean the agent is safe.
  • It does not replace a penetration test.
  • Areas it cannot review are marked as “not reviewable” in the report.

Frequently asked questions

Why a security check for agents?

Working code is not secure code. In a benchmark study (ICML 2026, 186 tasks), 57% of a coding agent's solutions were functionally correct, but only 11.8% were secure. The study shows the problem; it does not prove the effect of our review.

57%functionally correct
11.8%secure
Which agents are supported?

Today Claude Code, Claude Cowork and Codex CLI. In Cowork, instructions and skills are currently reviewed; permissions and network follow. You connect Codex CLI (version 0.147 or later) with the steps on the start page of the service. Gemini CLI is in preparation.

I use ChatGPT Dots. What can I do?

For ChatGPT Dots the portal offers a short check by questionnaire: 13 questions about the rights of your dot, free within the limits of your account. The result rests on your answers, not technically checked.

What does it cost?

The free account includes up to 3 agents, 5 checks per day and at most 75 per month, and stores reports for 30 days. No payment details, no contract term. The paid plans Starter (€19.90 net per month) and Team (€59 net per month) are planned and not yet bookable. All limits are listed under Plans and limits.

Why a company email address?

The account is meant for teams. You receive a code by email and your free account is created automatically with the first sign-in.

Does A2A Check change our agent on its own?

No. A2A Check delivers findings, suggestions and patterns and never changes anything about the agent itself. You or your agent make the changes in its workspace, after your approval.

I have questions first.

Write to us via the contact form. We reply as soon as possible.

About us

Who is behind A2A Check

A2A Check is developed and operated in Herne, Germany. Two founders are responsible for product, engineering and operations.

A2A Check, owner Mesut Yilmaz, Hauptstraße 202, 44649 Herne, Germany

Know what your agent may do, before it does it.

Your first review report in a few minutes. Free, no payment details.

Check for free

Sign in with your company email and a 6-digit code.