Become a sponsor to Pradyoth P

I work at the intersection of AI and security — red teaming production LLM systems before they reach users, then building the controls that keep them in scope. My day job is Gen AI red teaming at top firm.

My nights and weekends go into open-source security tooling for the same problem: how do you actually test whether an AI system is safe, instead of trusting a vendor's eval sheet.

I built Prompt Injection CTF, a gamified platform that teaches adversarial thinking against LLMs across 5 levels, mapped to OWASP's LLM Top 10 and MITRE ATLAS. I built AgentInjectionBench, the first benchmark targeting injection attacks in agentic, tool-using pipelines — the class of attack that single-turn benchmarks like AdvBench and HarmBench don't reach. I built Weighted Safety Refusal, a severity-weighted robustness harness on top of UK AISI's inspect_ai, because relative Z-score benchmarks hide the failures that actually matter.

None of this pays for itself.

Sponsorship covers the compute for running these evals at scale, the time to maintain them as models and attack surfaces shift, and the next tool in the pipeline.

If you're building with LLMs or agents and want the person red teaming them to keep testing them in public, this is where that happens.