Rez0s Sick Caching Bug & Local AI vs Subsidized tokens (Ep.191)
Source: YouTube · Critical Thinking - Bug Bounty Podcast · published Sep 10, 2026 · 37:32
[BLUF] The episode highlights the transformative power of AI in bug bounty hunting, featuring a complex prompt injection exploit discovered with Codex and a strategic shift toward local AI hardware investment.
Key Takeaways:
• AI-Driven Vulnerability Discovery: A hunter demonstrated a sophisticated bug involving an AI chatbot, using Codex to find a cache-based exfiltration vector that allowed data theft with a single user click 2:37.
• Prompt Injection Mechanics: The exploit utilized auto-invoking prompts and markdown image rendering to bypass Content Security Policies (CSP), leaking secrets index-by-index via server-side cache hits 4:06.
• Local AI as Infrastructure: Hunters are increasingly investing in local AI hardware (like Mac Studios and NVIDIA DGX Spines) to ensure redundancy, bypass token limits, and handle sensitive data that cloud models cannot process 14:24.
• Cost-Efficiency Strategies: To manage costs, some hunters use local models for data synthesis and report writing, reserving expensive cloud API calls for high-skill tasks like complex exploit development 19:30.
• Collaborative Hunting Models: The discussion touches on a new live hacking event format where the bounty pool is split among participants, fostering a more collaborative and less stressful environment 33:00.
The episode concludes that while cloud AI remains powerful, local infrastructure is becoming essential for professional hunters to maintain productivity and security. The community is also exploring new tools and formats to enhance collaboration and efficiency in bug bounties.
Sources:
- 2:37 Introduction of the "phenomenal" bug involving AI chatbot exploitation.
- 4:06 Explanation of the cache exfiltration technique using markdown images.
- 14:24 Discussion on purchasing high-end local AI hardware for redundancy.
- 19:30 Strategy for using local models to reduce cloud API costs.
- 33:00 Overview of the collaborative live hacking event format.
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: Penetration Testing. Commonly maps to: Security Assessment and Testing, 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
Did you say you bunty? >> You bunty. You say you bunty? >> No, Ubuntu. There's no you there's no Y at the beginning. >> What is this? What is this garbage? >> Okay, we're going to vote on this. That you're just definitely wrong. No one pronounces it with a Y at the beginning. >> I'm going to be roasted by the whole community. >> [music] >> Best possible Viking when you can just, you know, critical thing. >> [music] [laughter] >> What's up, guys? This week in bug bounty, we have one cool thing from Yes, We Hack. They have released a how to use Codex for bug bounty like course kind of blog write-up and it's actually like extremely verbose. It's got a lot of really cool data in it. So, a [snorts] lot of really cool things you can learn from. So, the one, you know, the question here they're tr…