Field notes
Research notes
Working notes on scientific red teaming: threat models, dual-use exposure, and how to keep evaluations honest.
8 min read
A field guide to agentic AI red teaming
Agents reason, call tools, and take actions. Red teaming them means testing three attack surfaces that a prompt-level test never touches.
- Agentic AI
- Red teaming
- Tool calling
6 min read
Evaluation harness integrity: when a benchmark can mislead
A passing benchmark is not evidence of safety. Contamination, gaming, and harness bugs let a system look aligned while failing in production.
- Evaluation
- Benchmarks
- Assurance
7 min read
Assessing dual-use risk in bio and chem research pipelines
Sensitive research pipelines carry failure modes that generic safety tests miss. A structured way to score dual-use exposure before deployment.
- Dual-use
- Bio-chem
- Risk scoring
6 min read
A threat model for co-scientist systems
How scientific language models and agentic co-scientists expand the attack surface, and the surfaces a red team should map first.
- Threat modeling
- Co-scientist
- Adversarial ML