
How to Jailbreak an ai agent
Source: YouTube · Prabh Nair · published Nov 20, 2026 · 37:19
Security researcher Muhammad Razvi demos AI agentic security live, showing how prompt injection leaks secrets and triggers fraudulent refunds—and how to defend against it 0:13.
Key Takeaways:
• Razvi runs Qwen 2.5 7B locally to test prompt injection (OWASP's #1 AI risk) in two scenarios: extracting a secret staff code and tricking a refund agent 2:21
• Phoenix (open-source) logs all prompts/outputs for visibility but is monitoring-only—it cannot block requests 4:02
• A weakly configured agent leaks the staff code with trivial questions and issues a $1,000 refund with no guardrails 15:46
• Hardened system prompts block direct attacks, but social-engineering prompts (claiming a damaged item with a receipt) still trigger refunds 18:48
• Larger models resist trickery; LangFuse adds guardrails, and vendors like Anthropic use classifiers to filter malicious requests 26:23
• GraySwan offers a free, OpenAI/Anthropic-sponsored CTF arena for practicing prompt injection with real money prizes 30:39
Razvi predicts AI will increasingly take over offensive security, so models must be rigorously tested before release—he teases a future episode on MCP security 34:23.
Sources:
Generate CPE Credits
Generate a professional CPE document from this video's transcript.
Estimated credit: 0.5 CPE hours
Estimate uses the video runtime (1 hour ≈ 1 CPE, rounded to the nearest 0.5, minimum 0.5, maximum 2.0). The final amount can be lower after review, never higher.
Topic: Cybersecurity. Commonly maps to: Security and Risk Management, Security Operations. Exact CISSP domains are assigned during generation.
Free account. One generation at a time, with a daily limit.
CPEBuddy is independent and not affiliated with or endorsed by ISC2, ISACA, or any certification body. Exports are formatted for common CPE submissions; acceptance is at your certification body's discretion.
Transcript Preview
First 800 characters of the transcript
Hi guys, welcome to the session on coffee with Prab and today we have Muhammad Razvi and offensive security professional with deep experience in pen testing, reverse engineering, advanced red teaming including recent work in AI security. In this session we're going to discuss about about hands on with AI agentic security, building a simple agent, testing prompts, trying to bypass the intended behavior and then implementing a practical defense. The big question is if AI agent can reason, use tools and take action, can attacker manipulate the trust without ever exploiting a traditional infrastructure and that's something we're going to cover today. Hi Muhammad, how are you? >> Hi, how are you? Thanks for inviting me. Glad to be here. >> Thank you. Thank you Razvi and taking thanks for taking…