I break AI systems for a living, identifying the line of whats possible becuase chances are next week it wont be. Currently im working on a startup trying to build LLMs that don't immediately fall over when pliny says "dance for me" and that empower users with a single dynamic and adaptive interface to commune with our soon to be token overlords.
Right now: Working on a personal AI project that I hope will raise some eyebrows.
Most Recently: AI Research & Product Development at Nous Research, where I spent my nights:
- Building LLM applications that empower users to work alongside LLM's
- Researching why AI security is harder than everyone pretends it is
- Actively trying to break llm's and discover there limits and plan for whats coming next
Before this: Spent years in traditional cybersecurity at VitalEdge, Varonis, and Juniper Networks learning how enterprise security actually works (spoiler: it barely does).
I'm interested in the gap between academic AI security research (cool in papers, useless in production) and what actually protects you when your AI is running in the wild getting poked by every script kiddie with a ChatGPT account.
I blog about AI security fails and what actually works at kellogg.pro/blog. Recent hits include:
- "AI Security Evolution: From Party Tricks to Actual Problems" - How we went from "lol I fooled an image classifier" to "someone just poisoned a model used by millions"
- "LLM Security Patterns: What Actually Works vs What Sounds Good in Meetings" - Battle-tested patterns for when your CISO asks "but is it secure?"
Less marketing fluff, more "here's what broke and why."
Languages: Python (for ML stuff), TypeScript/React (for frontends that don't suck), Go (when Python is too slow)
AI/ML: TensorFlow, PyTorch, whatever gets the job done. More interested in making models secure than making them 0.1% more accurate.
Security: Threat modeling, pen testing, the boring stuff that actually prevents breaches
Infrastructure: Kubernetes, cloud things, enough DevOps to be dangerous
Intelligent Security Platform - ML-powered threat detection that actually works in production (94% accuracy, deployed across 15+ companies, dramatically fewer "the sky is falling" alerts at 3am)
This Website - Built my own portfolio site because apparently I enjoy pain. Full-stack React app with an admin system that lets me write blog posts without touching code. GitHub-powered, Cloudflare-deployed, over-engineered in all the right ways.
Security isn't about preventing all attacks—it's about making attacks expensive enough that attackers go bother someone else. Perfect security is a myth. The goal is being "secure enough" that when everything inevitably breaks, you're not the one explaining to Congress why you stored passwords in plaintext.
For AI specifically: These systems are fundamentally gullible. They're trained to be helpful, and attackers exploit that helpfulness. Plan accordingly.
- Website: kellogg.pro
- Email: jon@kellogg.pro
- LinkedIn: linkedin.com/in/kelloggjon
Always interested in talking to people who understand that "it works in the demo" and "it works in production" are completely different things. Especially if you have stories about AI systems doing hilariously wrong things in the wild.
"Always carry a towel."