Office Hours: March 4, 2026

[andrewmchang]

Our next office hours are happening Wednesday, March 4, 2026 at 1700–1800 UTC!

Learn how T&S peers are using AI for safety workflows, ask for support in implementing open safety models, and connect directly with RMC model partners. We'll also discuss how you can contribute to ongoing projects that the RMC is working on!

• Discord event
• Google Meet
• #roost-model-community

Proposed Agenda

Introduction

• Recap of what the ROOST Model Community is and the purpose of office hours
• Updates from the RMC
• Personal intros

Demo of Phoebe

• How to automate safety operations by analyzing network threats and creating rules to detect and resolve emerging issues

Q&A / Show and Tell

• What questions do you have about the RMC? What resources would be most beneficial?
• What progress have you made in testing, using, or evaluating different open safety models?
• What use cases are you most interested in these models for? What would you like to see more of?-

Admin:

• Office hours Scheduling
• Contributing to upcoming RMC projects

What's Next:

• Introduce yourself in GitHub Discussions
• Start a new discussion! Whether it's an implementation question, research findings, or model specific feedback

[cassidyjames]

Notes

RMC's immediate objectives

1. Refine and publish eligibility criteria for models/partners
2. Expanding the models we work with/invite into RMC
3. Project sprints: specific things to work towards and produce within the RMC

Phoebe demo by Hailey

• It would be cool for this to be built-in somehow to something like the Osprey UI itself
• UX-wise, probably more guiderails rather than just a chat UI; e.g. a Claude skill or a button you hit to "investigate this account and draft a rule to address it"

Potential RMC projects

• Guide for how to implement open safety models in a T&S stack
  • How to not break the bank using the wrong models; e.g. how to use a general classifier that's really cheap to run, then send off to specialized models or humans for more review
• Library of example product policies across domains (e.g. abuse, spam, hate speech)
• Draft a short landscape analysis of major safety challenges unaddressed by existing RMC or open safety models

Community questions

• What types of risk should we test models against, and what are highest priority?
• How do you define licenses e.g. whether they're open source or not?
• What do we mean when we talk about "safeguarding" (within the community/ROOST)? AI safety, broader safety…?
  • Typically broader user safety: protecting users from harms like CSAM, grooming, abuse, spam, etc.
  • Also: the gpt-oss-safeguard model sometimes gets shortened to "safeguard," which can be confusing 🙈
  • More broad than "AI safety" even if there's some overlap; more "AI for safety" than "safety for AI"

Admin

Timing? 30 mins standing call, bi-weekly, starting 30 mins later (9:30 AM PST, 10:30 AM MST, 12:30 PM EST). Can always plan a longer session if we know we have a demo.

[andrewmchang]

Also: Interest in resourcing for taxonomy of risk categories!

[turt2live]

I wasn't able to make this office hours, sorry! Aside from the cost issue of running models, I'm particularly interested in what integration means for different models. I have limited experience in running models, and there's a few that I can't really figure out how to convert a huggingface link to something my code can call. Having that level of documentation would be helpful (a "how to accidentally break the bank" guide, if you will 😅 )

[andrewmchang]

Super helpful, thank you @turt2live! Definitely something we're interested in having at both a macro level and a model specific level