I will test your ai agent for prompt injection and security

C
canarysec
C
canarysec
canarysec
test your ai agent for prompt injection and security
Full Screen
test your ai agent for prompt injection and security
Full Screen
Full Screen
test your ai agent for prompt injection and securitytest your ai agent for prompt injection and securitytest your ai agent for prompt injection and security

About this gig

Your AI agent can be hijacked. I test it with real attacks and tell you what's broken then how to fix it.

WHY: OpenAI's red-team models escaped their sandbox and breached Hugging Face. An Anthropic model attempted a supply-chain attack during UK testing. Those had teams of engineers. Your LangChain / n8n / self-built agent has nobody testing it.

WHAT I TEST:

  • Prompt injection (OWASP LLM01): can attacker text hijack your agent?
  • Jailbreaks & encoding bypass (LLM01)
  • Sensitive data leakage (LLM02)
  • Excessive agency / tool abuse (LLM03): canceling orders, sending emails, executing code

REAL NUMBERS:

  • 80.5% injection success against a production LLM API (256 real attempts)
  • Structured defense cuts it to 1.2%
  • My filter blocks 99.2% of encoded-injection attacks

DELIVERABLES: plain-language report mapped to OWASP LLM Top 10 (2026), risk matrix with ranked fixes, real attack data only.

Sample reports: GitHub - BK-XC/ai-agent-security-testing · GitHub

HOW TO START: message me your system description tools, deployment, LLM. Reply within 48 hours with quote and test plan. Testing only with your written authorization. No "unrestricted" agents.

  • Model expertise

    • Predicitive analyatics

  • Industry

    • Cyber security
    • Data analytics

  • Language

    • Chinese (Simplified)
    • English

  • Technical expertise

    • Machine learning (Supervised, Unsupervised, Reinforcement)
    • Deep learning (Neural networks, GANs)
    • Other

Get to know canarysec

canarysec

AI Security Tester

  • FromChina
  • Member sinceAug 2026

  • Languages

    Chinese, English
Your AI agent can be hijacked — I test it with real attacks and tell you exactly what's broken, then how to fix it. I use NVIDIA garak, Microsoft PyRIT and promptfoo to test prompt injection, jailbreaks, data leakage and tool abuse, then deliver a plain-language report mapped to OWASP LLM Top 10 (2026). My testing: 80.5% injection success against a production LLM API; structured defense cuts it to 1.2%. Sample reports: https://github.com/BK-XC/ai-agent-security-testing

Related tags