The Work You Can't Check
Some work you can check in seconds. Some you can never check at all. That gap, verifiability, quietly decides how work gets trusted and priced, and where AI actually creates value.
ai verifiability security-operations
Alankrit Chona
Builder & researcher · verifiable AI for security
Some work you can check in seconds. Defending a company sits at the other end: nobody can prove a dismissed alert was benign. Whether work can be checked decides how far you can trust AI with it, and that's the problem I work on. I led the Cyber Defense Benchmark, which scores whether AI agents can find known attacks hidden in raw logs. I also build AI agents for security operations as co-founder and CTO of Simbian.
Research
The Cyber Defense Benchmark
Alankrit Chona (lead author), Igor Kozlov, Ambuj Kumar · arXiv, April 2026
It's the first public benchmark for open-ended threat hunting by AI agents. The agent gets raw Windows event logs recorded from real attack techniques, with no alert and no guiding questions, and has to find the malicious events on its own. Passing takes 50% recall on every ATT&CK tactic. No model has passed yet.
You can't prove a log is clean, but you can score whether an agent finds the attacks you know are in it. That's how the benchmark turns hunting into work you can check.
Start here
Three essays that build the argument: which work can be checked, why that favors attackers, and how to make defense measurable.
How to Train Your AI Attacker
Walk through the actual RL loop that makes an AI better at breaking in, cell by cell. Then run the identical loop for defense and watch it die at one cell, the reward, which is why offense keeps pulling ahead.
ai security-operations verifiability machine-learning
The Unbounded Investigation
One risky sign-in alert, five good analysts, five different investigations, all defensible. Why a security investigation has no natural end, why 'benign' has no proof, and the one move that makes detection measurable.
security-operations soc verifiability detection-engineering
What I write about
AI Systems & Agents
How agents work, where they break, and what it takes to trust them with work nobody can easily check.
Security & Verification
Why so much of defense resists measurement, and how to test a defense instead of trusting it.