Fighting Fire with Venom: Adversarial Defense Against...
summary
Venom is an inline reverse proxy (implemented on Cloudflare Workers) that fingerprints requests—headers, TLS fingerprints, IP clusters, request cadence—and routes high‑risk crawlers to a poison engine that injects tiny text flips or semi‑transparent image overlays while leaving user views untouched. A validation literature shows 0.1–2% poisoned training data can cut QA accuracy ~35% or provoke hallucinations; a mid‑2025 scan found ~9% of sites disallowing GPT‑bot and Cloudflare already blocks many crawlers by default. Venom will be open‑sourced but raises CFAA and ethical questions.
AI-written summary · about
chapters (11)
- 0:00Shadow copy of the internet
- 1:13AI data gold rush costs
- 3:44Voluntary crawling rules failed
- 7:16Websites hard gatekeeping
- 9:23Booby trap era with Venom
- 12:54Venom’s Drake Shake effect
- 14:44Accountability with micro weapons
- 17:54Flipping crawler’s cost-benefit
- 19:42Legal risks of data poisoning
- 21:09Venom for honest conversation
- 21:51Call for cooperation
related talks
43mOne Poisoned Artifact Can Steer Your AI: How Robust Are Your LLM-Assisted Security Workflows?
45mWinning by deception: tactics, techniques, and procedures of adversarial A.I.
42mHow to attack AI systems (and how to defend them) !!!!
48mCatch me if you can: Building a Web Malware Analyzer using Machine Learning
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
49mI Know Your Filtering Policy Better than You Do
44mDigital Cat-and-Mouse: Strategies to Outsmart Scrapers, Phishers, and Thieves
33mPoisoning Web-Scale Training Datasets is Practical