AI Security

AI/LLM Security Testing

Security assessment of LLM-powered applications, RAG pipelines, and AI agents covering prompt injection (direct and indirect), jailbreaks, RAG poisoning, training data extraction, model abuse, output handling, and supply chain risks. Aligned with OWASP Top 10 for LLM Applications.

Prompt InjectionRAG SecurityJailbreak & Model AbuseOutput Handling & PIIAI Supply Chain
Scope-based quote+ taxes
Contact usAll services
Process
6
phases
AI
4
tools
You get
7
deliverables

How it runs

  1. 01

    Model Inventory & Threat Modelling

    Catalogue the models, prompts, retrievers, agents, tools, and data sources in use. Build a threat model that maps trust boundaries, untrusted inputs, and the blast radius of each LLM action.

  2. 02

    Prompt Injection Testing

    Probe for direct and indirect prompt injection across user inputs, retrieved content, tool outputs, and uploaded files. Test system prompt leakage, instruction override, and persona escape.

  3. 03

    RAG & Retrieval Security

    Assess the retrieval pipeline for data poisoning, embedding manipulation, cross-tenant retrieval leakage, and index-level access control flaws. Test with crafted documents and adversarial queries.

  4. 04

    Output Handling & PII

    Test how downstream consumers handle model output: XSS via rendered markdown, SSRF via tool calls, code execution risks, and leakage of PII or training data through extraction prompts.

  5. 05

    Guardrails & Abuse Review

    Evaluate input filters, output classifiers, rate limiting, and abuse controls. Test cost-amplification attacks, denial of wallet, and agent loops that drive runaway resource consumption.

  6. 06

    Reporting & Recommendations

    Deliver findings mapped to OWASP LLM Top 10 with reproducible prompts, risk scoring, and concrete mitigations across prompts, architecture, guardrails, and monitoring.

AI assist

ai-toolkit.sh
AI-Assisted
$ cat tools.list
01
Adversarial Prompt GenerationGenerate diverse jailbreak and injection payloads beyond known public lists
02
RAG Poisoning SimulatorModel how poisoned documents propagate through retrievers and into model output
03
Output Classifier ProbingStress-test guardrails and content filters with edge-case prompts and encodings
04
PII Leakage DetectionIdentify when models surface training data, secrets, or cross-tenant content in responses
$ _

What you receive

  • AI/LLM threat model and architecture review
  • Findings mapped to OWASP LLM Top 10
  • Reproducible prompt injection and jailbreak proofs
  • RAG and retrieval security assessment
  • Output handling and PII leakage findings
  • Guardrail and abuse-control recommendations
  • Retest of critical findings after remediation

Ready to scope this engagement?

Every engagement is scoped individually. Get a tailored quote within 24 hours.

Request a Quote
AI/LLM Security TestingContact us