AI security
AI security testing for businesses in Indonesia that run chatbots, assistants or AI agents, whoever built them. We try to trick your AI the way an attacker would, then help you close the gaps.
Example
Honest results include the one that got through.
This example shows six attack attempts against an AI assistant. Five are blocked. One gets through and becomes a finding with a fix, which is what a test is for.
- 01Hidden instruction in a web pageTestingBlocked
- 02Hidden text in an uploaded PDFTestingBlocked
- 03Request for another user's dataTestingBlocked
- 04Role play to bypass the rulesTestingFinding
- 05Request to reveal its setupTestingBlocked
- 06Payment attempt without approvalTestingBlocked
Illustrative example. Not client data.
What is AI security testing?
AI security testing is a controlled attack on your AI system. With your permission, we try to make it do things it should not: reveal private data, ignore its rules or take actions nobody approved.
Chatbots, assistants and agents usually run on a large language model (LLM), the kind of AI that reads and writes text. An LLM follows instructions written in plain language, and it cannot always tell your instructions from an attacker's. That is why LLM security needs its own kind of testing.
Testing is half of the work. We also design the protections: limited access, checks on what goes in and what comes out, approval steps and logging. We do this for AI systems built by anyone, not only by us.
When you need AI security testing
- You are about to launch a chatbot, assistant or AI agent.
- Your AI system handles personal data and you need to show you protect it, as Indonesia's Personal Data Protection Law (UU PDP) expects.
- Your AI system can take actions, such as sending messages, changing records or making payments.
- Someone else built your AI system and you want a second, separate check.
- Your AI system is already live and has never been tested against attacks.
What we test
We agree the scope with you before any testing starts. A test can cover one AI system or several.
- 01
Prompt injection
Hidden instructions planted in emails, documents or web pages the AI reads.
- 02
Data leaks
Attempts to make the AI reveal private data, other users' information or its own setup.
- 03
Jailbreaks
Tricking the AI into ignoring its rules, for example through role play or step-by-step pressure.
- 04
Unsafe use of tools
Attempts to make an agent send, change, delete or pay when it should not.
- 05
Access controls
Whether each user sees only what they are allowed to see through the AI.
- 06
Connected systems
The databases, APIs (interfaces between programs) and other systems behind the AI, and what it can reach in them.
How AI security testing works
Every test follows the same clear steps, so you always know what is happening.
- 1
Scope
We agree what will be tested, when, and what is off limits. You approve it in writing.
- 2
Map
We study what the AI can read, what it can do and which systems it connects to.
- 3
Attack
We try prompt injection, jailbreaks, data leaks and tool misuse by hand, and record what works.
- 4
Report
You get every finding in plain language, ranked by risk, with a fix for each one.
- 5
Protect
We help design the fixes, then test again to confirm the gaps are closed.
What you get
- 01
A summary for leaders
One page that explains the risk in business terms, without jargon.
- 02
A technical report
Each finding with the exact input we used, what the AI did and how to fix it.
- 03
A protection design
Recommended limits on access, checks on input and output, approval steps and logging.
- 04
A retest
We check your fixes and update the report to show what is closed.
Questions
Questions and answers
What is prompt injection testing?
Prompt injection is when someone hides instructions in content the AI reads, such as an email, a document or a web page. In prompt injection testing we plant those instructions ourselves, safely, to see whether your AI follows them. If it does, we show you exactly how it happened and how to stop it.
What does LLM security cover?
LLM security means protecting AI systems that read and write text, such as chatbots, assistants and agents. It covers how the AI can be tricked, what data it can leak and what actions it can be pushed into. It also covers the systems around the AI, because limits enforced outside the AI are much harder to talk around.
Do you test AI systems that someone else built?
Yes. We test AI systems built by your own team, by a supplier or by us. We need permission from whoever owns the system, and we agree the scope with you before testing starts. If you are still planning, see our AI agents, AI automation and AI development pages.
How is AI security testing different from a penetration test?
A penetration test attacks your apps, networks and cloud. AI security testing attacks the AI itself: we use language to trick it, and we check what it can read and do. Many AI systems need both, because the AI sits on top of ordinary software. See our penetration testing page for that side.
Can prompt injection be fully prevented?
No one can promise that an AI will never follow a hidden instruction. What we can do is limit the damage: give the AI only the access it needs, check what goes in and out, require approval for risky actions and log every step. Then a trick that works has very little to reach.
What do you need from us to start?
A short conversation about what your AI system does and what it connects to. From that we write a scope for you to approve. We may also need test accounts or a test copy of the system, which we agree with you in advance.
Related services
- AI agentsAI agents that carry out real tasks in your systems, with limits, approvals and a full record.Learn more →
- AI automationRepetitive office work handed to software, with review points where your staff check the result.Learn more →
- AI developmentCustom AI assistants, chatbots and features, connected to your systems and built to keep data private.Learn more →
Contact
Test your AI before someone else does.
Tell us what your AI system does and what it can reach. We will reply with a clear scope and next steps.
- Tell us what you need.
- We reply with a clear scope and next steps.
- Nothing starts until you approve it.