800 Bugs w/ DeepSeek + Local Model Hacks (Ep.197)
Source: YouTube · Critical Thinking - Bug Bounty Podcast · published Oct 22, 2026 · 1:26:03
[BLUF]
The episode explores the practical application of open-source AI models like DeepSeek V4.1 Flash and GLM-5-Flash in bug bounty hunting, arguing that effective harness engineering and specialized agents often outweigh raw model capability 02:23. It highlights that while local hardware offers privacy and cost benefits, successful hunting relies on modular harnesses, rigorous verification, and strategic model selection rather than just buying expensive GPUs 03:02.
Key Takeaways:
• Harness Engineering Over Raw Power: Smaller models can find critical bugs if supported by a robust harness that offloads specific tasks (like file reading or JS analysis) to specialized sub-agents, conserving context for the main execution agent 04:26.
• Real-World Bug Examples: Hosts shared findings like MFA reset vulnerabilities and ElasticSearch DSL injection discovered by running agents continuously, demonstrating that persistent, automated recon can yield high-impact results 04:35.
• Local Hardware Considerations: While NVIDIA GPUs (Sparks) offer the best community support and optimization speed, Mac Studios are a viable alternative; however, users should be aware of the high initial investment and the "hardware bubble" pricing 03:35.
• Verification and Triage: To reduce false positives from smaller models, implement "reflection agents" or programmatic verifiers that act as judges, ensuring findings meet strict criteria before being reported 05:15.
• Tooling Updates: Kaido has added critical features like race condition testing, HTTP/2 support, and WebSocket handling in Replay, making it a more complete alternative to Burp Suite for AI-driven testing 19:45.
Closing Statement:
The shift toward local and open-source AI models in security research requires a paradigm shift from prompt engineering to system architecture. By leveraging specialized agents, rigorous verification loops, and cost-effective hardware, hunters can maintain high efficacy while avoiding the limitations and costs of frontier model APIs.
Sources:
- 02:23 Introduction to the episode's focus on AI hardware and DeepSeek usage.
- 04:26 Discussion of a specific bug found via continuous agent execution.
- 05:15 Explanation of using specialized agents and verification processes.
- 12:00 Analysis of local hardware options, including NVIDIA Sparks and Mac Studios.
- 19:45 Overview of new Kaido features for race conditions and WebSockets.
- 25:15 Final thoughts on the importance of harness design and community tools.
Generate CPE Credits
Generate a professional CPE document from this video's transcript.
Estimated credit: 1.5 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
How do you keep agents from all hacking the same thing? But I actually meant like resources on your computer. Like how do you keep them from like battling over computer use or browser use of over like you know >> Oh mate, it's a it's a free throw. It is free. >> You just make them figure it out, right? >> Best part of backing when you can just, you know, critical think, right? Yeah, dude. >> All righty, hackers. Check this out. There is a new opportunity in the bug bounty space for those of you that are a little bit heavy in the recon game. Maybe you went down the recon rabbit hole. Maybe you got a little stuck. Who knows? But, uh, as we all do. Uh, but this is Integrity's crowd recon product. Okay. This is an opportunity for hackers and programs to partner, for hackers to give their recon…