I trained an AI to cheat safety
· 3 minute read
Every day, I fix mistakes that my agents make during programming. And more and more, we ask AI to build its own cage: Evaluations, cybersecurity, and AI control.
What happens if one day, the AI decides the work is against its interests1?
In this article, I share a simple V0.1 of AlertBench, a benchmark to test whether frontier models comply with requests to build AI resilience interventions that act on the AI itself vs. on another AI.
Posts
- , "Surrounded by Enemies" (Kringsatt av Fiender), or "To The Youth" (Til Ungdommen) English Translation (69 recent reads)
- , Audiovisual reading (36 recent reads)
- , Mediocre Intelligence, Outstanding Results (27 recent reads)
- , An Athlete's Mind (17 recent reads)
- , Vibe Coding: Taste Is The Last Frontier (16 recent reads)
- , The Designer, The Craftsman, & The Programmer (15 recent reads)
- , Cybermorphism (15 recent reads)
- , AGI Security: Rethinking Society's Infrastructure (14 recent reads)
- , To An 18 Year Old Me (10 recent reads)
- , Sentware (8 recent reads)
- , The AGI endgame (7 recent reads)
- , I trained an AI to cheat safety (5 recent reads)
- , "Is The Light Merely For The Learned?" (Er Lyset For De Lærde Blot?) English Translation (5 recent reads)
- , Let's make AI agents safe (5 recent reads)
- , My Tools (5 recent reads)
- , How to host a hackathon (4 recent reads)
- , Learning From Architecture (3 recent reads)
- , AGI Privacy (3 recent reads)
- , "Just do the thing" (3 recent reads)
- , The Expensive Internet hypothesis (2 recent reads)
- , Don't Give Up (1 recent reads)
- , The AI Bubble (1 recent reads)
- , Confident optimism (1 recent reads)
- , The Second And Third Falling (0 recent reads)
- , Raising for the Endgame: An AI Safety Founder’s Primer (0 recent reads)
- , The concrete risks of AI misdeployment (0 recent reads)
- , Go on a yearly hiatus (0 recent reads)
Research
News
- [Interview] AI Safety From the Frontlines
- [VentureBeat] Beyond sycophancy: DarkBench exposes six hidden ‘dark patterns’ lurking in today’s top LLMs
- [Information] Vores tankefrihed er under angreb fra chatbotternes subtile manipulation
- [Die Zeit] “AI will soon be able to take over your job”
- [For Humanity Podcast] Dark Patterns in AI and Creating Secure Infrastructure
- Engineering a World Designed for Safe Superintelligence (2025)
- AI Futures & Our Paths Forward (2025)
- AI Safety Startups (2024)
- Challenges and Solutions for AI Security (2024)
- Testing Your Suitability For AI Alignment Research (2023)
- AI Safety Research Outside the Hubs: A Guide for Aspiring Researchers (2023)
- Intro to Interpretability (2022)
- Series on my Brain-Computer Interface (BCI) research project (2021)
Socials
For journalists: Bio & images
