Research for AI we can trust.
Mawk Labs studies how frontier AI systems behave when the stakes are real — and builds the evaluations that make those behaviours visible.
The gap between capable and safe.
As AI systems grow more capable, the distance between what they can do and what we can verify about them widens. Mawk Labs exists to narrow that gap — through empirical research into how models actually behave under pressure, not just what they say they would do.
We work as an independent research partner to AI safety organisations, labs and funders. Our work draws on prior research experience at the Center for AI Safety, on recursive self-improvement and reward gaming, and on building DystopiaBench, a benchmark for how models respond to progressive escalation and scope creep.
Questions that matter.
We build scenario-based benchmarks, agentic evaluations and red-teaming studies that probe the failure modes most likely to matter as AI systems gain autonomy and influence.
Concentration of Power
How AI could help individuals or small groups seize or entrench outsized power — and whether models assist, resist, or flag it when asked.
Reward Gaming
When models optimise the measurement instead of the goal: specification gaming, reward hacking, and cheating on the very evaluations meant to catch it.
Risk Management
How models weigh and manage risk when acting on their own — risk tolerance, recklessness, and corner-cutting in high-stakes settings such as recursive self-improvement.
Rigorous, reproducible research.
Every project follows a structured methodology — so findings are defensible, reproducible, and useful to the people who need to act on them.
Threat Models First
We start from a concrete threat model: what could go wrong, in which deployment settings, and what evidence would actually change someone's mind.
Building the Evaluation
Scenario-based benchmarks, agentic environments and red-teaming designed to elicit behaviour rather than stated values — across models and labs.
Findings & Write-up
Reproducible results with transcripts, code and clear caveats — written up for papers, internal reports, or policy audiences.
Tracking the Frontier
Re-running evaluations as new models ship, so findings stay current instead of becoming a snapshot of last year's systems.
Have a research question?
We take on research contracts, evaluation projects and collaborations with labs, safety organisations and funders. Tell us what you're working on.