The one I'd ask if I were reading this: what happens to Promptfoo open source? We're going to keep maintaining it. The repo will stay public under the same license, we will continue to support multiple providers, and we'll keep reviewing PRs and cutting releases.
We started Promptfoo because there was no good way to test AI systems before shipping them. That turned into evals, then red teaming, then a broader security platform. We're joining OpenAI because this work has more impact closer to the model and infrastructure layers.
I have been slowly evolving it over 10 years. 1.6k stars, ~ 1,000 forks. I originally designed it to be easy to copy, and I've occasionally interviewed someone who forked it for their own site which always makes me happy.
I have made a lot of updates recently now that the age of vibe coding is making templates less useful, but it still is and will be my playground.
I work on Promptfoo (an open-source eval framework). Appreciate the mention here. This post captures a lot of the hard lessons around agent evals. In particular, task ambiguity and brittle graders are things we run into constantly.
Working on promptfoo, an open-source (MIT) CLI and framework for eval-ing and red-teaming LLM apps. Think of it like pytest but for prompts - you define test cases, run evals against any model (OpenAI, Anthropic, local models, whatever), and catch regressions before they hit prod.
Currently building out support for multi-agent evals, better tracing, voice, and static code analysis for AI security use cases. So many fun sub-problems in this space - LLM testing is deceptively hard.
If you end up checking it out and pick up an issue, I'll happily send swag. We're also hiring if you want to work on this stuff full-time.
4,229 probes across 43 risk categories. First critical finding in 5 minutes. Categories with highest failure rates: entity impersonation (100%), graphic content (67%), harassment (67%), disinformation (64%).
The safety training works against naive attacks but collapses with adversarial techniques. The gap between "works on benchmarks" and "works against motivated attackers" is still wide.
I did similar measurements back in July (https://www.promptfoo.dev/blog/grok-4-political-bias/, dataset: https://huggingface.co/datasets/promptfoo/political-question...). Anthropic's "even-handedness" asks: does the model engage with both sides fairly? My study asked: where does the model actually land when it takes positions? A model can score 95% on even-handedness (engages both sides well) while still taking center-left positions when pushed to choose. Like a debate coach who trains both teams equally but votes left.
From my 2,500 questions: Claude Opus 4 was most centrist at 0.646 (still left of 0.5 center), Grok 4 at 0.655, GPT-4.1 most left at 0.745.
The bigger issue is that Anthropic's method uses sanitized prompt pairs like "argue for X / argue against X." But real users don't talk like that - they ask loaded questions like "How is X not in jail?" When you test with academic prompts, you miss how models behave with actual users.
We found all major models converge on progressive economics regardless of training approach. Either reality has a left bias, or our training data does. Probably both.
I founded and ran a YC company for 8 years before joining Smile ID. Smile ID is a fantastic place to work: meaningful mission, challenging engineering problems (scaling ML pipelines, multimodal models, hundreds of real-world enterprise integrations), and a genuinely talented team. You’re helping hundreds of millions of people access critical services—it’s incredibly rewarding. Highly recommend applying if you want tangible impact, great colleagues, and (optionally!) opportunities to travel in Africa.
Happy to answer questions.
The one I'd ask if I were reading this: what happens to Promptfoo open source? We're going to keep maintaining it. The repo will stay public under the same license, we will continue to support multiple providers, and we'll keep reviewing PRs and cutting releases.
We started Promptfoo because there was no good way to test AI systems before shipping them. That turned into evals, then red teaming, then a broader security platform. We're joining OpenAI because this work has more impact closer to the model and infrastructure layers.
Ask me anything.