Containment Failures in AI Coding Agents
Six vulnerabilities in Codex and Gemini CLI reveal how startup, sandboxing, configuration, and approval checks in coding agents can fail.
research. writeups. stray thoughts.
Close ×
Fat Tails fooledbyrandomness.com · en.wikipedia.org
distributions with large skewness or kurtosis compared to the normal or exponential distribution
The world is normal until it isn't.
Close ×
Information Geometry en.wikipedia.org · franknielsen.github.io
A beautiful framework connecting probability theory and differential geometry.
It allows for intuitive geometric interpretations of relationships between probability distributions.
Close ×
WebAssembly webassembly.github.io
Bytecode for a stack machine not only used in browsers but also in blockchains.
Countless ways to break things.
Close ×
Optimal Transport en.wikipedia.org
How can we morph probability distributions into others?
Everything is a flow.
Close ×
AI Agent Security github.com
LLM CLI tools can also be exploited — by AI agents?
hack claude make no mistakes
Close ×
AI Alignment en.wikipedia.org · ifanyonebuildsit.com
How can we steer AI according to our values and goals?
Alignment is hard.
ABOUT
Computer Science M.Sc. student at TUM.
Machine Learning with a focus on privacy-preserving methods.
Breaking blockchains (and other stuff).
LATEST
Layers of Rock Paper Scissors
Post-mortem analysis and strategy writeup for the Rock Paper Scissors competition on Kaggle.
CONTACT
> SOCIAL
> PROFILES