Luis Cosio
Hardening frontier AI systems against nation-state adversaries and loss-of-control failure modes.
I work at the intersection of frontier AI and high-security systems. My focus is turning abstract safety and security requirements into deployable technical artifacts that can withstand real adversaries, from nation-state attacks to loss-of-control scenarios.
Focus Areas
- Frontier AI Security (SL5): Securing model weights, inference systems, and lab infrastructure against nation-state threats and loss-of-control failure modes.
- AI Alignment and Interpretability: Mechanistic interpretability, scalable oversight, and monitoring-based interventions.
- Security Engineering: Air-gapped deployments, attack surface minimization, and high-assurance runtime design.
Blog
October 9, 2026
SIEVE
How an AI data center campus is set up, how it sheds heat, and how to detect it.
May 23, 2026
Distillation Caltrops
Watermarks that live in learned habits, not in output tokens.
January 25, 2026
μInference: From Minimal Stack to SL5 Weight Enclave
Radically minimized inference with seL4-backed weight protection.
January 25, 2026
OpenWild
A new conversational dataset for modern language models.
January 15, 2025
Mechanistic Watchdog
Real-time cognitive interdiction for LLMs.
All posts →
Selected Work
- Security Level 5 Task Force - SL5 security artifacts for frontier AI systems
- MechWatch - 1st place, Defensive Acceleration Hackathon (writeup)
- MATS Mentor - Mentoring scholars in the ML Alignment & Theory Scholars Program
- SPAR Mentor - Mentoring researchers starting careers in AI safety and alignment
- Algoverse Mentor - Guiding student research teams in AI safety
Background
Mexico City based technologist with 15 years across cloud, cybersecurity, and AI. Currently working full-time on AI safety and security.