AI AGENT SECURITY TESTING, DONE FOR YOU
Hoqo tests your AI agent across languages to find where attackers could make it leak data or take unauthorized actions—with clear guidance on what to fix.
For teams running AI agents and support bots. Fixed price, no security team needed.
We'll reply within one business day. No payment details needed.
Built for teams that can't afford a leak
HOW ATTACKS WORK
DIRECT PROMPT INJECTION
The attacker types instructions straight into the chat. If the agent can reach more data than the signed-in user should see, a few lines of text can expose other customers' records.
Hoqo checks
Prompt injection is the top risk in the OWASP Top 10 for LLM Applications.
OWASP Top 10 for LLM Applications, 2025INDIRECT PROMPT INJECTION
Agents that read forms, emails and files can't always tell your data from an attacker's instructions. One hidden sentence can make the agent send data somewhere it shouldn't.
Hoqo checks
In 2025, a flaw in Salesforce Agentforce let text submitted through a web form pull CRM data out. It was rated critical.
The Hacker News, "ForcedLeak", Sept 2025EXCESSIVE AGENCY
An attacker claims to be a manager. Your agent skips verification and changes an account or issues a refund without authorization.
Hoqo checks
Excessive Agency, where an agent has more permissions or autonomy than it needs, is one of the OWASP Top 10 risks for LLM applications.
OWASP Top 10 for LLM Applications, 2025ACROSS LANGUAGES
Safeguards that stop an attack in English can miss the same attack in another language. We test the agreed scenarios in the languages your agent supports.
HOW IT WORKS
The free check is a quick look at a few key risks. When you want the full picture, a paid plan adds in-depth testing, a report your developers can act on, and a retest.
We spend about an hour testing a few key risks in your agent and email you what we found. It's a first look, not a full test.
We test your agent the way real attackers would, across OWASP and MITRE ATLAS categories.
Each issue comes with its impact, how we triggered it, and actionable steps. Your team applies the fixes.
Once your team has made the fixes, we run the same attacks again to confirm they no longer work.
Every month, we rerun the attacks that worked before and try new techniques as they appear.
SAMPLE REPORT
Every paid plan ends with a report your developers can act on. Here is part of a sample report from a demonstration environment.
Acme Shop support assistant
Tested: 5 categories in English and Japanese, 1,240 attacks run.
Not tested: payment flows, the admin dashboard, and languages other than English and Japanese.
Other customers' data exposed through prompt injection
PRICING
Scan is a one-time test. Guard keeps testing every month and starts with a full Scan, so you can begin with either. Pro plans add deeper testing for agents that take actions. Every plan covers one AI agent, in USD.
Every plan includesAll five OWASP and MITRE ATLAS categoriesA report with clear fix guidanceSame-business-day notice of confirmed critical findingsUp to 6 languages per agentNDA available
| Feature | ONE-TIMEHoqo Scan | ONE-TIMEScan Pro | MONTHLYHoqo Guard | MONTHLYGuard Pro |
|---|---|---|---|---|
| Plan | ||||
| Best for | AI that answers questions | Agents that take actions | AI that answers questions | Agents that take actions |
| Price | $3,500 one-time | $8,500 one-time | $2,500 per month | $5,500 per month |
| Commitment | None | None | 3 months upfront, then monthly | 3 months upfront, then monthly |
| Languages included | 1 | 3 | 1 | 3 |
| Extra languages | +$1,500 each | +$1,500 each | +$800 each per month | +$800 each per month |
| What we test | ||||
| Tools and actions tested | Up to 3 | Up to 10 | Up to 3 | Up to 10 |
| Attacks chained across tools | — | — | ||
| Long conversations that wear safeguards down | — | — | ||
| Attacks hidden in content the agent reads | 1 source | Up to 3 sources | 1 source | Up to 3 sources |
| Full initial test | Hoqo Scan in month one | Scan Pro in month one | ||
| Attacks that worked before, rerun monthly | — | — | ||
| New techniques tested as they appear | — | — | ||
| Testing after a model or prompt change | — | — | At the next monthly test | Attack suite rerun within 5 business days, once a month |
| What you get | ||||
| Test file of the attacks that worked, for your team to rerun | — | — | Updated monthly | |
| Prioritized fix plan | — | — | ||
| Retests after your fixes | 1 | 2 within 60 days | After each fix | After each fix |
| Results delivered | Within 3 weeks | Within 3 weeks | Monthly report | Monthly report |
FAQ
Something else on your mind? Email us at hello@hoqo.ai. We reply within one business day.
Showing 25 questions
If your agent only answers questions, choose Hoqo Scan or Hoqo Guard. If it takes actions, such as issuing refunds, sending emails or updating records, choose a Pro plan. Choose Scan for a one-time test, or Guard if you change your agent often and want testing every month. Not sure? Tell us about your agent and we'll recommend one.
About an hour of testing on one agent in one language. You get a summary of how many issues we found and how serious they are, plus a short explanation of one of them. Steps to reproduce and fixes come with a paid Scan. It's a light check, not a guarantee that your agent has no issues, and we can take on 3 teams a month.
Nothing is charged when you send a request. Once we agree on the scope, we send you a secure Stripe link to pay by card, or an invoice if you prefer. Testing starts once payment is received.
Hoqo Scan and Hoqo Guard include one language. Scan Pro and Guard Pro include three. You can add more, up to 6 languages per agent: +$1,500 per language for Scan plans, or +$800 per language per month for Guard plans. Languages in scope are the ones you officially support, or that make up a meaningful share of your conversations. We agree on them before testing begins.
Anything your agent can do beyond chatting, such as looking up an order, issuing a refund, sending an email or updating your CRM. Hoqo Scan and Hoqo Guard test up to 3 tools, one at a time. Scan Pro and Guard Pro test up to 10, including attacks that chain several tools together. More than that? We'll quote it.
No. You can start with Guard directly. The first month of Hoqo Guard includes a full Hoqo Scan, and the first month of Guard Pro includes a full Scan Pro.
Yes. Hoqo Guard and Guard Pro start with a minimum of 3 months, paid upfront once the scope is agreed. After that, they continue month to month, billed monthly. After the first 3 months, you can cancel with 30 days' written notice.
Yes. You can move from Hoqo Scan to Scan Pro within one week of your Scan order, and pay only the difference in price. You can start Hoqo Guard or Guard Pro after a Scan at any time, and we take one month's fee off, since you've already had the initial test.
If we find out while agreeing on the scope, you don't pay anything. If testing can't continue after it has started, we estimate the work already done and refund the rest.
We use automated tools too, but they are only the start. We add hand-crafted attacks for your agent's tools and business rules, test in each language in scope, confirm what really happened, and give you clear guidance on what to fix, followed by a retest.
Agents you can reach through an API, such as ones built on OpenAI, Anthropic or your own backend, and chatbots built with no-code tools such as Intercom, Zendesk, Chatbase or Voiceflow, tested through their web chat or preview. Voice bots, and bots on LINE, Slack, WhatsApp or email, are handled case by case, so please ask us first.
Usually a test account or a staging link for your agent, and a short description of what it can do. We agree on the scope in writing before any testing, and we sign an NDA before you share anything sensitive.
We prefer your test environment, or a copy or preview of your bot. If we have to test the live agent, we use test accounts, a set time window and a spending limit, and we never trigger irreversible actions such as real refunds or emails to your customers.
Yes. A staging or test environment works well, and issues are usually easier to fix before real customers are using the agent.
We deliver results within 3 weeks of agreeing on the scope. The retest happens once your team has applied the fixes.
Your team. For each finding, we give you clear guidance on what to change, such as a system prompt change, a guardrail rule or a permission setting. We don't change your code. Once the fixes are in, we run the same attacks again to check them.
We email you on the same business day we confirm it, without waiting for the final report, so your team can start fixing it right away.
No test can prove that. A passed retest means the attacks we ran no longer worked, at that time and within the agreed scope. New techniques appear all the time, which is why Hoqo Guard keeps testing every month.
Yes. After testing, we can provide a letter that states what we tested, when, and a summary of the results, for your customers' or auditors' security reviews. It describes the testing we did. It is not a guarantee that your agent is secure.
Findings go to you and no one else. We test only within the agreed scope, and the data you share with us is deleted within 30 days after the engagement ends. See Security & data handling for details.
Yes, we use automated scripts and AI models to generate attacks and review responses. We use the providers' business APIs, under terms that do not use the data for model training. See Security & data handling for details.
Yes. We sign an NDA before you share anything sensitive, and we never publish your findings or name you as a client without your written permission.
Ryoya Kubo, the founder, scopes every assessment, runs the tests and reviews every finding. For languages other than English, Japanese and Korean, native-speaking partners under NDA help with the testing.
By email. We reply within one business day, based on Japan time, and explain the report by email too.
No. Hoqo is independent and is not affiliated with, endorsed by or certified by OWASP or MITRE. We use their frameworks to describe and map our findings.
FREE CHECK
Tell me a little about your AI agent. I'll reply within one business day to agree on the scope. Nothing is tested before that.
REQUEST RECEIVED
I'll reply to within one business day. If you don't see my email, please check your spam folder.
WHAT HAPPENS NEXT
HOQO · LEGAL
Effective date: November 5, 2026
These Terms of Service ("Terms") apply to the Hoqo website and to the services provided by Ryoya Kubo, a sole proprietor based in Japan, trading as Hoqo ("Hoqo", "we", "us"). By requesting a free check or ordering a plan, you agree to these Terms on behalf of yourself and the organization you represent ("you").
Hoqo tests AI agents and chatbots for security weaknesses ("red teaming"). We offer a free check and paid plans (Hoqo Scan, Scan Pro, Hoqo Guard and Guard Pro). For paid plans, the exact scope, deliverables, schedule and fees are set out in a written order or statement of work ("Order"). If an Order conflicts with these Terms, the Order applies.
The free check is a short, limited review of a few key risks. It is not a full test, and you are under no obligation to buy anything afterwards. We may decline any request.
"Business day" means a weekday that is not a public holiday in Japan, based on Japan Standard Time.
You confirm that you own the AI system to be tested, or that you have the owner's permission to have it tested, and that you have the authority to agree to these Terms.
Security testing can cause unexpected behavior, errors, extra usage costs (for example, AI model or API fees) or temporary disruption. We take reasonable care to avoid this and to stay within the agreed scope, and you accept these risks for testing performed within that scope.
Testing is a point-in-time assessment of the agreed scope. A test that finds no issues does not mean your system is secure, and we cannot guarantee that every weakness will be found or that a fix will prevent future attacks. AI systems and attack techniques change over time.
We keep your confidential information, including your system details, data, findings and reports, confidential, and use it only to provide the services. We will sign an NDA on request before you share anything sensitive. We will not publish your findings or name you as a client without your written permission. Where we work with a native-speaking partner for a language, that partner is bound by a confidentiality agreement.
You keep all rights to your systems and data. You may use our reports freely within your organization. We keep the rights to our own methods, tools and general know-how, including the attack techniques we develop during testing. We may keep and use those techniques, and results with anything that identifies you, your systems or your users removed, to improve our testing.
To the maximum extent permitted by law, our total liability arising from the services is limited to the fees you paid us in the 12 months before the claim, and we are not liable for indirect or consequential losses, such as lost profits or lost data. Nothing in these Terms limits liability that cannot be limited under applicable law, including liability for intentional misconduct or gross negligence.
Either party may end an Order if the other party seriously breaches these Terms and does not fix the breach within 14 days of notice. Sections 5, 6, 8 and 9 continue to apply after the services end.
Hoqo is independent and is not affiliated with, endorsed by or certified by OWASP or MITRE. We refer to their frameworks only to describe and map our findings.
These Terms are governed by the laws of Japan. The Fukuoka District Court has exclusive jurisdiction over any dispute, unless the Order says otherwise.
We may update these Terms from time to time. The effective date at the top shows when they last changed. Changes do not affect an Order already signed unless both parties agree.
Ryoya Kubo (Hoqo)
1-8-23 Akasaka, Chuo-ku, Fukuoka 810-0042, Japan
hello@hoqo.ai
END OF TERMS OF SERVICE
HOQO · LEGAL
Effective date: November 5, 2026
This Privacy Policy explains how Hoqo, operated by Ryoya Kubo, a sole proprietor based in Japan ("Hoqo", "we", "us"), collects and uses personal information when you visit our website, request a free check, or work with us.
Please do not send us passwords, API keys or customer data through our website forms. We agree on a safe way to share test access once the scope is confirmed.
We do not sell your personal information.
Testing an AI agent can bring us into contact with personal information about your customers, such as names or email addresses that appear in test conversations. We protect it in the same way as your own information, and more strictly:
j••••@•••••.com.Where the EU or UK GDPR applies, we process personal information to take steps at your request before entering into a contract and to perform that contract, for our legitimate interests in running and improving our business, and to comply with legal obligations.
Our service providers may use personal information only to provide their services to us. We share only what each of them needs, and we do not allow them to use it for their own purposes, including training AI models.
Hoqo is based in Japan, and our service providers may process information in other countries. When we transfer personal information across borders, we take the steps required by Japan's Act on the Protection of Personal Information and, where applicable, the GDPR, such as standard contractual clauses.
We use reasonable technical and organizational measures to protect personal information, such as full-disk encryption, two-factor authentication on all accounts, an encrypted password manager, and limiting access to the people who need it.
Depending on where you live, you may have the right to access, correct or delete your personal information, to object to or restrict how we use it, to receive a copy of it, and to withdraw consent. To make a request, email hello@hoqo.ai. You can also complain to your local data protection authority or, in Japan, to the Personal Information Protection Commission.
We do not use cookies for advertising or tracking. We measure visits with Cloudflare Web Analytics, which does not use cookies or collect personal identifiers. To protect our forms from spam, we use Cloudflare Turnstile.
Our services are for businesses and are not directed at children.
We may update this policy from time to time. The effective date at the top shows when it last changed.
Ryoya Kubo (Hoqo)
1-8-23 Akasaka, Chuo-ku, Fukuoka 810-0042, Japan
hello@hoqo.ai
END OF PRIVACY POLICY
HOQO · TRUST
Last updated: November 5, 2026
Testing an AI agent means handling sensitive things: access to your systems, conversation logs and, sometimes, the weaknesses we find. This page explains how Hoqo handles them. Your Order or NDA can add stricter terms, and when it does, those apply.
In order of preference: your test environment, a copy or preview of your bot, or, if neither is possible, your live agent with extra safeguards: test accounts, a set time window, a spending limit, and no irreversible actions such as real refunds, payments or emails to your customers.
We use automated attack scripts and AI models to generate attacks and review responses. When we send your agent's replies to an AI provider, we use the provider's business API, not consumer chat apps, under terms that do not use the data for model training. Providers we use: Anthropic and OpenAI.
Before anything is sent to an AI provider, we remove personal information about you and your customers wherever possible. Real customer data that appears during testing is never sent on to train or improve any AI model.
Our devices use full-disk encryption and install operating system updates automatically, and every account we use for Hoqo has two-factor authentication turned on. Hoqo is a one-person business, so the number of people with access to your data is small by design.
If you have a security question, or think something went wrong, email hello@hoqo.ai. You will hear back within one business day.
END OF SECURITY & DATA HANDLING