AI & LLM Security TestingAI agents and chatbots that cannot be talked into leaking data.
Security testing for LLM apps, AI agents and chatbots. We test for prompt injection, data leaks and unsafe actions, then add the guardrails that stop them.
- OWASP Top 10 for LLMs
- Prompt injection testing
- Agent guardrails
- WhatsApp and voice bots
- Prompt injection blocked · hidden instruction in a PDF
- System prompt no longer revealed
- Agent refund limit enforced
- Customer data masked in logs
AI agents follow instructions. Including the wrong ones.
An LLM can be tricked by text hidden in an email, a web page or a document. When that agent can read data or take actions, the risk is real. We test for it before launch.
Direct and indirect injection tested
Untested and assumed safe
Least-privilege access per user and task
Agent can read everything
Approvals and limits on refunds, emails and updates
Agent acts without limits
Prompts, keys and personal data protected
System prompt and records exposed
Logged conversations with alerts on misuse
No record of what the AI did
Data flows mapped for DPDP and client reviews
Personal data sent anywhere
In-house penetration testers with OSCP and CEH, reviewed by an ISO 27001 Lead Auditor.
Scope, rules of engagement and NDA signed before we touch a system.
Independently audited information security.
One person who knows your systems, from scoping to retest.
Why AI systems need security testing
Chatbots and AI agents take instructions in plain language, and they cannot always tell your instructions apart from text an attacker has slipped into an email, a web page, a support ticket or a PDF. When the same agent can search your documents, update a CRM or ERP record, issue a refund or send a WhatsApp message, a successful attack becomes a real incident.
AI and LLM security testing finds these weaknesses before customers or attackers do. We use the OWASP Top 10 for LLM Applications as the baseline.
What we test
- Direct prompt injection and jailbreak attempts from users
- Indirect prompt injection through documents, emails, web pages and tool results
- Leakage of system prompts, API keys, customer records and other users’ data
- Excessive agency: what the agent can do, and whether limits and approvals hold
- Retrieval (RAG) access control, so users only see documents they are allowed to see
- Logging, monitoring and abuse limits
Guardrails we put in place
Testing is only half the job. We can add input and output filtering, least-privilege access for each tool the agent uses, human approval for sensitive actions such as refunds or bulk messages, rate limits, masking of personal data in logs, and alerts when the system is misused. We also document what personal data the system processes and where it goes, which supports your obligations under India’s DPDP Act. Final compliance decisions should involve your legal adviser.
How the engagement works
- Scoping. We review the use case, the model and platform, the data sources and the actions the AI can take.
- Red-team testing with a library of attack prompts plus scenarios written for your business.
- Report with each finding, how it was triggered, its impact and the recommended guardrail.
- Fixes, by your team or ours, followed by a retest.
Sigosoft builds AI agents, WhatsApp and voice assistants and workflow automation, so we test with a builder’s understanding of how these systems are put together. Testing is delivered by our in-house security team with specialist partners, under ISO 27001:2022 certified processes. See also AI agents and workflow automation and voice and WhatsApp AI.
What we secure.
AI chatbots
Website, WhatsApp and voice assistants.
AI agents
Agents that read email, update CRM or ERP and send messages.
RAG systems
Document search and knowledge bases built on your data.
LLM integrations
OpenAI, Anthropic, Gemini and open-source models.
Guardrails
Input and output filters, approvals and rate limits.
AI governance
Usage policies, logging and data handling for audits.
Scoped in days. Report in two weeks. Retest included.
- 1Day 1–3
Scope
What to test, test accounts, dates and rules of engagement, under NDA.
- 2Week 1–2
Test
Automated coverage, then manual testing by certified testers. Critical issues reported the same day.
- 3End of week 2
Report & fix
Verified findings ranked by risk, with fixes for your stack. We can make the fixes.
- 4Within 60 days
Retest
Every fix retested and confirmed, with an updated report for auditors and clients.
AI systems we have built.
“Third agency we tried. First one that actually delivered.”Khalid Al-Rashid · Head of Operations, StratEdge GroupView all projects
Prompt injection is when text supplied by a user, or hidden in a document, email or web page, changes what an AI model does, for example making it ignore its rules or reveal data. It is the first risk in the OWASP Top 10 for LLM Applications.
No. We test chatbots and agents built by any team or vendor, as long as you own them or have permission to test them.
We test applications built on OpenAI, Anthropic, Google Gemini and open-source models, including RAG systems and agents built with n8n, Make or custom code.
Yes. We build AI agents ourselves, so we can add guardrails, limit permissions, add human approval steps and set up logging for you.
It helps. We map what personal data your AI system collects, where it is sent and how long it is kept, which supports your obligations under the DPDP Act. Final compliance decisions should involve your legal adviser.
Launch your AI agent with the risks tested, not guessed.
Tell us what you need tested or certified. We reply with a scope, timeline and cost within 2 hours.

