Applications are open for our first cohort — apply by June 15, 2026.
Phanguard

Independent ML-safety research

Understanding AI systems from the inside out.

Phanguard is a small, independent research project working on mechanistic interpretability and ML safety. We publish open research and run a free program for students entering the field.

Fiscally sponsored by Hack Club (501(c)(3)) Open research Free student program
Read our research Apply to the program →

About

We study the internal structure of neural networks to make AI systems safer and more transparent.

Phanguard is an independent research project, fiscally sponsored by Hack Club (The Hack Foundation, a registered 501(c)(3)). We're a small founding team doing open work in ML safety and interpretability — currently focused on sparse autoencoders, and we run a free research program for students entering the field.

Focus
Sparse autoencoders (SAEs)
Areas
Interpretability · ML safety
Program
Free · 12 students per cohort
Fiscal sponsor
Hack Club (501(c)(3))

Research

Current work

Our research centers on understanding how neural networks internally represent knowledge, and using that understanding to build safer systems.

Mechanistic interpretability

Active

Investigating how neural networks form internal representations, such as through the circuits, features, and computational structures that drive model behavior. The aim is to make otherwise opaque models legible to researchers and auditors.

Circuits · Feature analysis · Transparency

Reasoning via sparse autoencoders

In progress

We're co-authoring an in-progress paper exploring how sparse autoencoder architectures can extract and analyze the features underlying reasoning in language models. SAEs decompose activations into interpretable directions; we're applying this to study chain-of-thought reasoning.

SAEs · Reasoning · Language models

Program

A free research program

A structured, free research experience for high-school students.

Duration
3 months
Cohort size
12 students
Cost
Free
Eligibility
High-school students (13–18)

  1. Conduct original research

    Work on a real research question in ML safety, interpretability, or a related area and not coursework.

  2. Write for publication

    Draft a paper with the goal of submitting to a workshop or preprint server.

  3. Work with a mentor

    Each student is paired with a volunteer mentor for technical guidance and feedback throughout the cohort.

  4. Present your findings

    Share your work with the cohort and, where it's a good fit, get support submitting it to a workshop or preprint server.

Get involved

Apply

The program is free and open to high-school students (ages 13–18). We're accepting 12 students for our first cohort. No specific background is required. We look for genuine curiosity and commitment. Applications close June 15, 2026.

Students

Join the cohort as a student researcher. No prior research experience required.

Apply via Google Form →

Peer mentors

For high-schoolers and undergrads who want to support the mentorship team, help guide students, review code, and keep things collaborative.

Apply via Google Form →

Volunteer mentors

Experienced researchers and practitioners who can guide a student or two. The time commitment is flexible (a few hours a week).

Reach out on LinkedIn →

Details

Frequently asked

Is the program really free?

Yes. The program is free for everyone we accept. Costs are covered through donations and our fiscal sponsor.

Who is eligible to apply?

Our first cohort is for high-school students aged 13–18. A genuine interest in AI safety and interpretability matters more than prior experience.

Do I need prior research experience?

No prior publication experience is required. A solid foundation in programming (Python) and basic machine-learning concepts will help you make the most of the three months.