Joe Torres and Alex Kleinman

Executive Director, Active Site; Co-Founder, Active Site

  • Joe Torres is the Executive Director of Active Site and has worked on uplift studies, mirror biology countermeasures, broad-spectrum antivirals, and vaccines (Alvea). Prior to Active Site and Alvea, he was an intellectual property technology specialist at Clark & Ebling. He completed his PhD in Molecular and Cell Biology at UMass Amherst and has over 10 years of research experience in molecular biology and immunology.

    Alex Kleinman co-founded Active Site, a research nonprofit that generates empirical data from the physical world on biosecurity risks. His work spans wet lab research and human trials: he co-led the largest RCT on LLM-driven biological uplift in novices (2026; n = 153) and contributed to ABC-Bench (NeurIPS 2025), an agentic biosecurity benchmark. He also co-authored an eLife (2026) study on antibiotics and vaccines against mirror bacteria. Previously, he worked on broad-spectrum vaccines at Alvea.

  • We’re looking for candidates to work with on the following research topics or to suggest a topic of their own:

    Measuring Expert Uplift
    AI companies have identified expert uplift in biology as a critical risk. The core concern is that AI models could significantly enhance an expert's ability to develop biothreats worse than COVID-19.

    The Role: Help us rigorously measure if—and to what extent—AI systems (LLMs and biological design tools) enhance experts in operationalizing a safe, relevant proxy task.

    The Goal: Design and pilot a statistically robust, biologically sound human-subjects study to gather real-world data on this risk. Your experimental design must be safely executable in a BSL-2 lab within a reasonable time frame, be scalable, and capable of generating a highly powered, feasible sample size.

    Evaluating Agentic Lab Automation
    Lab automation hardware is rapidly advancing toward a future where capable AI agents could direct fully or nearly autonomous workflows, requiring minimal to no skilled human intervention.

    The Goal: Design and pilot physical world AI-directed lab automation evaluations that measure capabilities in autonomous bio.

    Uplift Bench
    Running randomized controlled trials takes time. What if we could use our data to build a digital benchmark that tells us when and what to run next?

    The Goal: Build a validated "synthetic RCT" or "video-game" benchmark that has real-world meaning by using our trial data or by collecting data of your own.

    OSINT Vulnerability Identification
    We rely on collaborators to surface the most salient threat models for evaluation, but we're increasingly moving to building our own systems for identifying and patching new vulnerabilities.

    The Goal: Build semi-automated systems for identifying and evaluating new AIxbio advances, synthetic biology capabilities, and biorisk proliferation.

  • We are seeking fellows who combine deep technical expertise with a strong bias toward action.

    Ideal candidates will have:

    Domain Expertise: A background in cell/molecular biology, virology, bacteriology, AI-bio design tools, lab automation, mechanical/electrical engineering, or benchmark development.
    AI Fluency: Familiarity and high proficiency in leveraging LLMs to accelerate both research and operational workflows.
    Experimental Mindset: A track record of designing and running rapid, MVP-style experiments to test hypotheses quickly.


    Work Style: We highly value researchers who are highly collaborative, exceptionally communicative, and capable of driving projects independently.