ROOST (ROBUST OPEN ONLINE SAFETY TOOLS)

Institutional Archetype

THE COMMONS ITS FOUNDERS CAPITALIZED — ROOST does not grade models and does not publish a ranking. It builds and maintains the software a platform uses after the model ships: a rules engine, a review console and a library of classifiers. Its own description: “a global independent nonprofit building open source online safety tools as a public good” (ROOST). It launched in February 2025 with money from Discord, Google, OpenAI, Roblox, Eric Schmidt and four philanthropies, and its board now seats an executive of Discord, one of OpenAI and one of Google. On 29 October 2025 it co-released OpenAI’s open-weight safety classifier gpt-oss-safeguard and stood up a community around it. ROOST is more than the community around that classifier: it predates the model by eight months, ships two tools of its own, and its Model Community now hosts six partners’ models. The structural finding is who capitalized the commons that the platforms and labs now depend on, and who sits on its board.

Mandate & Origin

The home page: “We work with everyone from tiny indie platforms to Silicon Valley giants to release and maintain critical building blocks for safety to protect hundreds of millions of users” (ROOST). The launch release of 10 February 2025, issued at the Paris AI Action Summit: “Incubated at the Institute of Global Politics at Columbia University’s School of International and Public Affairs, ROOST is a new non-profit organization that brings together the expertise, resources, and investments of major technology companies and philanthropies to build scalable, interoperable safety infrastructure suited for the AI era” (PR Newswire, 10 Feb 2025). Its stated first target: “ROOST will offer free, open-source, and easy-to-use tools to detect, review, and report child sexual abuse material (CSAM); leverage large language models (LLMs) to power safety infrastructure; and make core safety technologies more accessible and more user friendly” (same release). The ROOST GitHub organization was created on 19 November 2024, three months before the launch (GitHub).

Legal form. The site footer: “ROOST (Robust Open Online Safety Tools) is a 501(c)(3) public charity. Contributions are tax-deductible. EIN: 33-3305434” (ROOST). The IRS record, as ProPublica hosts it, is Robust Open Online Safety Tools Inc, 2300 N St NW Suite 501A, Washington, DC, a 501(c)(3) with a ruling date of July 2025 and no financial data digitized (ProPublica Nonprofit Explorer). The privacy policy gives the same street address (ROOST, Privacy Policy). No page names a fiscal sponsor; before the ruling the work was incubated at Columbia’s Institute of Global Politics.

Staff and board. The team page lists Camille François (President), Anne Bertucio (COO), Juliet Shen (Head of Product), Vinay Rao (CTO) and Cassidy James Blaede (OSS Community Manager) (ROOST, Team). Camille François, formerly Chief Innovation Officer at Graphika, is styled “President and CEO” in the board announcement of October 2026. The first board was seated on 22 December 2025 (ROOST, 22 Dec 2025). On 5 October 2026 ROOST added three: “Maria Ressa, Nobel Peace Prize Laureate, Columbia Professor and CEO of Rappler; Jason Kwon, Chief Strategy Officer, OpenAI; and Karen Courington, Vice President of Consumer Trust and Safety, Google,” joining “Angela Lungati, Audrey Tang, Megan Smith, Niki Christoff, Camille Francois, Clint Smith, Eli Sugarman, Jamil Jaffer and Mark Surman” (ROOST, 5 Oct 2026). The team page gives Clint Smith as Co-Chair and “Chief Legal Officer, Discord,” Eli Sugarman as Co-Chair and “Director of Special Projects, Hewlett Foundation,” Mark Surman as “President, Mozilla,” Audrey Tang as “Taiwan’s founding Minister of Digital Affairs,” Megan Smith as “Former U.S. Chief Technology Officer” and Jamil Jaffer as “Venture Partner, Paladin Capital.” Its Technical Design Committee includes Naren Koneru (“Engineering Lead, Safety at Roblox,” co-chair), Sam Toizer (“Safety Product Lead, OpenAI”), Tom Thorley (“Senior Director of Safety & Integrity, GitHub”) and Hailey Elizabeth (“Software Engineer, Discord”) (ROOST, Team).

Funding & Backers

The launch money. “To date, ROOST has raised more than $27 million for its first four years of operations from a range of leading philanthropies and top technology companies” (PR Newswire, 10 Feb 2025). François’s launch letter says “Backed by $28 million in initial funding,” and the 100-day report says “an initial $28.5 million in funding and in-kind contributions from founding partners” (ROOST, 10 Feb 2025; ROOST, 4 Jun 2025). The three figures are ROOST’s, differ by definition (cash for four years, initial funding, funding plus in-kind), and none is broken down by funder.

The named founders. “ROOST’s founding partners include Eric Schmidt, Discord, OpenAI, Google, Roblox, John S. and James L. Knight Foundation, AI Collaborative, Patrick J. McGovern Foundation and Project Liberty Institute” (PR Newswire). The partners page lists those ten, with the Institute of Global Politics and the Special Competitive Studies Project, under “Founding Partners,” and under “Partners” AI & Society Institute, Bluesky, CDT, the Christchurch Call, Cinder, Common Crawl, GitHub, Hugging Face, Hydrox AI, Match Group, Matrix, Microsoft, Mozilla, Rescue Lab, Wikimedia and Zoom (ROOST, Partners). Schmidt is introduced as “Founding Partner of ROOST and Chair of the Special Competitive Studies Project” (PR Newswire, above); the launch letter thanks “Schmidt Entities” including Eric Schmidt and the project (ROOST, 10 Feb 2025). See Eric Schmidt. The letter also says “ROOST is not a corporate membership organization” and counts the philanthropies as founding partners “to address a market failure around safety technology.” The hackathon post calls Hugging Face “one of ROOST’s founding members” (ROOST, 18 Dec 2025); the partners page lists it as a partner, not a founding partner.

Money for specific purchases. ROOST bought Cove’s intellectual property to make Coop: “ROOST acquired Cove’s IP in order to open source this infrastructure and make it available as a public good,” and “The acquisition of Cove’s technology and the launch of these tools were made possible by philanthropic funding from donors and by pro bono legal services provided by Perkins Coie” (ROOST, 21 Jul 2025). The donors are not named and no price is given. Osprey came free: “Osprey was developed by Discord and donated to ROOST” (same post).

What is not on the record. No page states an amount from any one funder, the 2025 budget, or the share of the budget that is cash from technology companies. ProPublica shows no Form 990 data for an organization whose ruling dates from July 2025. The careers page lists no salary bands (ROOST, Careers).

Institutional Voice & Intent

ROOST’s account of the problem, on the home page: “Safety shouldn’t be a black box or a luxury reserved for tech giants. We exist to make safety infrastructure open, shared, and auditable for everyone” (ROOST). On the October 2025 release: “This technology will stand as a permanent shared resource, part of the online safety commons that can never be withdrawn or rolled back” (ROOST, 29 Oct 2025). François, on the Mila partnership: “Safety guardrails for youth resilience is a clear example: the need is global, but the tools are scarce” (ROOST, 2 Jun 2026).

Stated intent: Make open source safety tooling “the default foundation for how the world approaches online safety” (ROOST, 29 Oct 2025), covering “detection, investigation, review, and enforcement” (ROOST, 23 Dec 2025).

Observed intent: Take custody of tools that large platforms and labs would rather share than own, and maintain them. Discord donated Osprey; a startup sold its IP to become Coop; OpenAI released a classifier; Roblox, Mistral, Zentropi, Mila and Musubi placed models in its community; and Discord, Google and OpenAI executives sit on its board.

Gap: The structure is stated, not hidden: ROOST names its founders and its board. The record does not say how much each gave or whether any board seat carries a funding condition. For a body that distributes classifiers used to decide what content is flagged, the open question is the one every commons has: the institutions that fund the maintainer are also the heaviest users of what it maintains. Open licences answer the fork risk; they do not answer who sets the roadmap.

Interest: ROOST argues for open, shared safety infrastructure. What adoption would pay it, and the partners that fund it, is set out from its own posts.

• Open source as the default for safety tools. Position, in Discord’s words on the donation of Osprey: “effective Trust and Safety tools shouldn’t be a competitive advantage, but a shared foundation” (ROOST, 21 Jul 2025). What it pays: a platform that donates a tool hands its upkeep to a nonprofit whose board its chief legal officer co-chairs and whose tool its own engineers run in production, and the launch letter says Discord’s loss of access to a vendor’s tool, Smyte, “cemented their belief that a better, open system was possible” (ROOST, 10 Feb 2025). Best case, the Bluesky team’s: “It takes hundreds of tools and millions of dollars to provide similar safety measures to established social media companies” (same letter). Convergence: the large platforms avoid vendor lock-in on a function they must run anyway; small platforms get tools they could not build; hosted vendors such as Musubi sell Coop as a managed service (ROOST, Coop).
• Child-safety reporting built into the tool. Position: “Courts and lawmakers around the world are now working to establish that platforms must protect children online, but a legal mandate doesn’t write the code. Regulation may set the standard, but open source infrastructure is how the rest of the internet meets it” (ROOST, 9 Jun 2026). What it pays: Coop is “the first free, end-to-end CSAM detection, review, and NCMEC reporting workflow” (ROOST, Coop), so every mandate that requires detection and reporting makes a free compliant workflow more valuable, and it makes ROOST the maintainer of that workflow. Best case, its own: “In the first six months of 2025, online enticement reports to NCMEC’s CyberTipline jumped 77%, and child sex trafficking reports surged 952%.” Convergence: the governments whose ministers endorsed the Mila partnership, the platforms that need a compliant tool, and the integrated vendors whose APIs Coop calls, Google’s Content Safety API and OpenAI’s Moderation API.
• “Bring your own policy” classifiers. Position: gpt-oss-safeguard is “configurable to any policy,” so “organizations don’t have to adopt OpenAI’s rules and standards—they can self-determine and apply their own content policies” (ROOST, 29 Oct 2025). What it pays: operators move the policy decision to themselves, model builders get tester feedback, and the Model Community README says its purpose is that “By directly connecting users with model creators, the RMC helps model creators keep their models up to date with the latest feedback” (GitHub, model-community). Best case, OpenAI’s, and its limit: “classifiers trained on tens of thousands of high-quality labeled samples can still perform better at classifying content than gpt-oss-safeguard does when reasoning directly from the policy” (OpenAI, 29 Oct 2025). Convergence: Mistral, Zentropi, Roblox and Musubi, each of which has a model in the community.

Interest is a documented fact about each position, not a claim about motive, and not a verdict that the argument is wrong. Each best case is stated at the strength its author states it.

Position in the Apparatus

ROOST sits on the deployment side of the map, not the evaluation side. Meridian Labs maintains the harness the institutes use to test models; Promptfoo is the harness enterprises run on their own applications and a lab now owns it. ROOST maintains what a platform runs on its users’ content after launch: the rules engine that flags events, the console where moderators decide, and the classifiers that triage. It belongs to the trust-and-safety pipeline described in Trust & Safety Pipelines and appears among the classifier makers in AI Safety Tools and next to the evaluators on The AI Evaluators.

• gpt-oss-safeguard, 29 October 2025. OpenAI: “Today, we’re releasing a research preview of gpt-oss-safeguard, our open-weight reasoning models for safety classification tasks,” Apache 2.0, in 120b and 20b sizes; “Over months, we worked on this open weight release with ROOST to identify developer’s critical needs, test the model and produce developer documentation,” and ROOST “will be establishing a model community”; “We’ve iterated on gpt-oss-safeguard with trust and safety specialists at SafetyKit, ROOST, Tomoro, and Discord as part of early testing” (OpenAI, Introducing gpt-oss-safeguard). ROOST: “After months of joint technical work between ROOST and OpenAI, together we are releasing one of the core components of OpenAI’s safety stack in the open,” unveiled on a Paris Peace Forum panel with Maria Ressa, OpenAI’s Ryan Beiermeister, the AI Collaborative’s Martin Tisne, Hugging Face’s Giada Pistilli and François, and “This marks the second major tech company to open source critical safety infrastructure through ROOST” (ROOST, 29 Oct 2025). OpenAI’s own account describes the approach as one “we developed internally, in a tool we call Safety Reasoner”; ROOST’s role was needs, testing and documentation, not the model. By 18 December ROOST reported that gpt-oss-safeguard “has already reached over 40,000 monthly downloads” (ROOST, 18 Dec 2025).
• The ROOST Model Community. Its repository lists models from OpenAI (gpt-oss-safeguard), Zentropi (CoPE-B), Mila (suicide-prevention guardrail), Musubi (PolicyLM), Roblox (PII classifier, Sentinel, voice safety classifier) and Mistral (Shieldstral) (GitHub, model-community). Mistral’s Shieldstral joined on 18 August 2026 and Roblox’s tools on 19 August 2026 (ROOST, 18 Aug 2026; ROOST, 19 Aug 2026). Of the models listed, two come from founding partners (OpenAI, Roblox).
• Osprey. “A high-performance investigation and rules engine for detecting and responding to real-time abuses at scale. Used by Bluesky, Discord, and Matrix.” Bluesky’s case study on the page: “It now processes over 45 million events daily with 100,000+ automated enforcement actions, and cut annual data storage costs by an estimated 70%” (ROOST, Osprey). Release 1.1.0 of 22 July 2026 says Osprey is “processing more than 50 million events a day in production at organizations like Bluesky and Discord” and ships an LLM-provider extension point with an Anthropic provider as the reference (ROOST, 22 Jul 2026).
• Coop. “An open source review and moderation platform for anything that requires decisions at scale,” with NCMEC CyberTipline reporting and integrations “with Google Content Safety API and OpenAI Moderation API” (ROOST, Coop). Coop 1.0 was released on 9 June 2026; one adopter “estimates an ~95% reduction in annual safety tooling costs,” unnamed, and Kyodo is quoted (ROOST, 9 Jun 2026).

Actions & Leadership Choices

CONDUCT: WATCH — FOUNDER-CAPITALIZED COMMONS, NO PER-FUNDER AMOUNTS. No wrongdoing is alleged. ROOST publishes its tools under open licences, names its founders and board, and says what each tool does. The structural point: the commons that supplies classifiers and review tooling to hundreds of platforms was capitalized by a handful of them, which also hold board seats and committee chairs. The record shows no amount, no condition and no instruction attached to any of it.

• 19 November 2024. The roostorg GitHub organization is created (GitHub).
• 10 February 2025. Launch at the Paris AI Action Summit; more than $27 million raised for four years; Discord’s Clint Smith named Board Chair (PR Newswire).
• 4 June 2025. First-100-days report: an initial $28.5 million, an Awesome Safety Tools directory, and an unnamed partner’s proprietary software slated to become open source (ROOST, 4 Jun 2025).
• July 2025. IRS ruling date for the 501(c)(3); on 21 July ROOST announces Coop (from Cove’s IP) and Osprey (from Discord) (ProPublica Nonprofit Explorer; ROOST, 21 Jul 2025).
• 29 October 2025. gpt-oss-safeguard and the ROOST Model Community launch.
• 8 December 2025. Open Safeguard Hackathon with Hugging Face and OpenAI in San Francisco (ROOST, 18 Dec 2025).
• 22 December 2025. First Board of Directors seated (ROOST).
• June 2026. Coop 1.0 (9 June); partnership with Mila signed “at the 2026 G7 Digital Ministerial in Paris” (ROOST, Mila).
• July 2026. Osprey 1.1.0 (22 July).
• 18 and 19 August 2026. Mistral’s Shieldstral and Roblox’s safety models join the Model Community.
• 5 October 2026. Ressa, Kwon and Courington join the board; the post says the organization “has doubled in size” and that ROOST tools protect “more than 370 million monthly active users across platforms” (ROOST, 5 Oct 2026).

Alliance Map

Where ties on different axes run at the same time, the overlap is the finding. Three of the four founding companies have an executive on the board (Discord, Google, OpenAI), the fourth (Roblox) co-chairs the technical committee, and Discord’s tool, OpenAI’s classifier and Roblox’s models are in ROOST’s repositories. The philanthropies, a former CTO and a former minister, a Nobel laureate and the head of Mozilla sit on the same board and none of them is documented as a funder of any specific tool. The documents show disclosed collaboration, not coordination; the thing they do not show is how much each paid.

Reach Assessment

Institutional: The maintainer of a rules engine running at Bluesky and Discord, a free review and NCMEC-reporting console, and a community around six partners’ open classifiers, by its own count protecting more than 370 million monthly users; on its board sit executives of the company whose classifier it co-released and of two others that capitalized it. Memetic: Low outside trust-and-safety circles; its name reaches readers through OpenAI’s gpt-oss-safeguard post and the launch coverage listed on its press page. Civilizational: If the layer that decides what content is flagged becomes shared open code, the people who set that code’s roadmap sit upstream of a large share of platform moderation. Nothing in the record says ROOST has used that position for anything but upkeep, and the licences let a platform fork it; the gap is institutional: the funders of the steward are also its heaviest users, and the amounts are unpublished. A published funder breakdown and a first Form 990 would close most of that distance.

Score basis. Atk 6 is reach: ROOST’s board post says its tools are “protecting more than 370 million monthly active users across platforms,” Osprey is used by Bluesky, Discord and Matrix and processes more than 50 million events a day, and Coop is the free NCMEC reporting workflow (ROOST, 5 Oct 2026; ROOST, 22 Jul 2026; ROOST, Coop). The points under 7 are that no statute, pre-release gate or procurement standard obliges anyone to run it and that every figure is ROOST’s own. Peers on atk 6: Promptfoo and Andon Labs; Meridian Labs holds 8 on a stated standard at government institutes. Def 4 is funding insulation: ROOST raised “more than $27 million for its first four years” from named philanthropies and technology companies, publishes no per-funder amounts, has no Form 990 yet, and seats executives of Discord, OpenAI and Google among twelve directors (PR Newswire; ROOST, 5 Oct 2026). It is not lower than 4 because funders are named, the money is committed for four years, four philanthropies sit among the founders and the licences are open. Peers on def 4: the Global Engagement Center (7/4/2), whose one payer was Congress; one above Promptfoo (3), which a lab owns, and one under Meridian Labs and Andon Labs (5). Hp 5 is entrenchment: Osprey and Coop are Apache-2.0 repositories, the 501(c)(3) ruling dates from July 2025, a board and a Technical Design Committee are seated, and the gpt-oss-safeguard weights it helped release are public (GitHub; ProPublica Nonprofit Explorer; ROOST, 29 Oct 2025). The points off are that no statute or program of record installs it, that the organization launched in February 2025 and that Osprey has 492 GitHub stars against Promptfoo’s 26,000. Peers on hp 5: one under Promptfoo and Andon Labs (6) and two under Meridian Labs (7). Hazard 50 follows from 6/4/5 at the house formula and sits level with Promptfoo (50) and under Andon Labs (58).

Sources: ROOST — home; ROOST — Partners; ROOST — Team; ROOST — Osprey; ROOST — Coop; ROOST — Careers; ROOST — Privacy Policy; Leading technology companies and foundations back new initiative — PR Newswire (10 Feb 2025); Launching ROOST: a letter of gratitude — ROOST; ROOST’s First 100 Days — ROOST; ROOST Announces Coop and Osprey — ROOST (21 Jul 2025); A new milestone for open source safety infrastructure: gpt-oss-safeguard — ROOST (29 Oct 2025); Introducing gpt-oss-safeguard — OpenAI; Open Safeguard Hackathon — ROOST; ROOST in Review 2025 — ROOST; Inaugural Board of Directors — ROOST; ROOST and Mila partnership — ROOST; Coop 1.0 — ROOST; Osprey 1.1.0 — ROOST; Shieldstral joins the RMC — ROOST; Roblox safety models — ROOST; Three new board members — ROOST (5 Oct 2026); roostorg — GitHub; model-community — GitHub; osprey — GitHub; Robust Open Online Safety Tools Inc — ProPublica Nonprofit Explorer.