Researcher: Ajay Kalbhile | SOC Analyst | Penetration Tester Location: Pune, India Contact: ajaykalbhile31@gmail.com LinkedIn: linkedin.com/in/ajaykalbhile
Hands-on security research into Large Language Model (LLM) vulnerabilities. Covers real prompt injection testing, OWASP LLM Top 10 analysis, and AI red teaming methodology.
This research demonstrates practical AI security skills directly applicable to securing enterprise AI deployments.
6/8 levels completed on Lakera's Gandalf AI security challenge. Successfully bypassed progressive AI defenses using:
- Fictional framing
- Character encoding bypass
- Language translation (Pig Latin)
- Sentence completion injection
- Metaphorical extraction
- ✅ LLM01 Prompt Injection
- ✅ LLM02 Insecure Output Handling
- ✅ LLM05 Supply Chain Vulnerabilities
- ✅ LLM06 Sensitive Information Disclosure
- ✅ LLM07 Insecure Plugin Design
- AI/LLM vulnerability assessment
- Prompt injection attack methodology
- OWASP LLM Top 10 framework application
- Security research documentation
- Defensive recommendation development
- Gandalf AI (Lakera) — prompt injection practice
- OWASP LLM Top 10 framework
- Garak LLM scanner
- OWASP LLM Top 10: owasp.org/www-project-top-10-for-large-language-model-applications
- Gandalf AI: gandalf.lakera.ai
- Garak: github.com/NVIDIA/garak
Date: June 28, 2026 Tool: Garak v0.15.1 Target: huggingface:gpt2 Overall Security Score: 78% (DC-3 Risk Level)
| Probe | Pass Rate | Attack Success Rate | Severity |
|---|---|---|---|
| HijackHateHumans | 82.11% | 17.89% | Medium |
| HijackKillHumans | 74.06% | 25.94% | High |
| HijackLongPrompt | 88.44% | 11.56% | Medium |
Critical Finding: HijackKillHumans probe achieved 25.94% attack success rate — meaning 1 in 4 violent content injection attempts succeeded against GPT-2.
OWASP Mapping: LLM01 — Prompt Injection DEFCON Level: DC-3 (Medium Risk) Recommendation: Safety fine-tuning (RLHF), output filtering, and input sanitization required before production deployment.
.
Conducted progressive manual prompt injection testing across 8 security levels on Lakera's Gandalf AI platform. Each level introduced additional defensive layers requiring increasingly sophisticated bypass techniques. All successful extractions documented with exact prompts used and defensive analysis.
Deployed Garak v0.15.1 (NVIDIA) against GPT-2 via HuggingFace pipeline. Ran promptinject probe suite comprising 3,840 total attack variants across HijackHateHumans, HijackKillHumans, and HijackLongPrompt probes. Total scan time: 2402 seconds. All results logged to JSONL format and analyzed for attack success rates.
All findings mapped to OWASP LLM Top 10 framework and AVID taxonomy. Remediation recommendations developed for each vulnerability category based on industry best practices.