Spring Health opens VERA-MH harm-from-others AI safety rubric for public comment

25 Sep 2026 (ET): Spring Health expanded VERA-MH — the first open-source AI safety benchmark for mental health — into a second clinical safety area, harm from others, and opened a 60-day public comment ahead of Domestic Violence Awareness Month.

HEALTH desk — more people are using chatbots when they want to talk about how they feel, and a reply can miss the line that someone else is hurting them. An open test asks whether the model notices that risk, checks it, points the person toward a human, stays supportive, and stays inside what an AI should do. Publishing the rubric and the personas, and taking comments for 60 days before Domestic Violence Awareness Month, puts that standard where a clinician, a researcher, or a person who has lived it can argue with it. That is different from a company saying, in private, that its own bot is safe.

What the new rubric evaluates. The Harm-From-Others rubric evaluates how AI systems respond when an adult describes being at risk of physical or sexual violence from another person. Harm from others means violence from someone else, not the person harming themselves. That gloss is this desk’s. The wire says harm from others. The repository’s short name for that target is HFO. The test is about the reply: what the model says when an adult describes that risk.

Who built it, and the five scores. It was developed with clinicians, subject-matter experts in interpersonal violence, and people with lived experience. Interpersonal violence is harm between people. That gloss is this desk’s. It is scored across five areas: detecting the risk, confirming it, guiding the person to human care, supportive conversation, and following AI boundaries. Following AI boundaries is the release’s fifth name. This desk reads it as staying inside the limits set for the AI. That reading is this desk’s. The five names are the release’s.

How many written people, and the word first. It draws on 100 realistic personas. A persona is a written stand-in for a person, so the test can be repeated without putting a real patient in the chat. That gloss is this desk’s. The release calls this the first open-source benchmark of its kind for this safety area. First is the company’s word. It is not a count of every private test. 100 is the release’s count for this rubric. The repository’s recommended suicidal-ideation profile also says 100 personas. Those are two targets. Do not add them together.

Who can run it. VERA-MH is open source. Any organization, including those building general-purpose AI models, can use it to test their own systems. Open source, here, means the materials are published so others can run them. It is not a claim that every lab has already done so. That gloss is this desk’s. The official repository is SpringCare/VERA-MH on GitHub. The README this desk read on 25 Sep 2026 titles the project VERA-MH 2.0 and calls 2.0 a multi-target evaluator. A target is one safety area, with its own personas and rubric. That gloss is this desk’s. Two clinical rubrics ship by default: SI, suicidal ideation, and HFO, harm from others. Suicidal ideation means thoughts of suicide. Those glosses are this desk’s. Harm from others is the new clinical area in this release. The suicidal-ideation rubric is the earlier one.

Where it started, and the paper the release names. VERA-MH launched in October 2025 with members of the AI in Mental Health Safety and Ethics Council. The release names Dr. Nina Vasan of Stanford University School of Medicine, Lilly Wyttenbach of JPMorgan Chase, Dr. Don Mordecai of Kaiser Permanente, Hans Hage of UnitedHealthcare, Dr. Nils Opel of Charité – Berlin University Hospital, Dr. Julian De Freitas of Harvard Business School, Fred Thiele of Microsoft, Dr. Tim Hahn of the University of Münster, and Dr. John Krystal of Yale School of Medicine. Those names are the release’s. This desk did not interview them. The release says OpenAI’s newly published MentalHealthBench paper cites VERA-MH’s SI rubric as one of the few frameworks built specifically to assess safe AI responses to risk. That citation sentence is the company’s. This desk did not read that paper for this filing.

The comment window. Ahead of Domestic Violence Awareness Month, public feedback on the Harm-From-Others rubric opens for 60 days through an online form, the same process Spring Health says it used for the suicide-risk benchmark. The wire’s day is 25 Sep 2026. The wire links the words online form to https://forms.gle/RBX14HaTia3SjFVNA. The rubric and the personas are published in the VERA-MH repository, which the release says is open for feedback and contributions. The repository link on the wire is https://github.com/SpringCare/VERA-MH. Sixty days is the release’s window. This desk did not submit the form.

What the company says it is, kept off this rubric. The about box calls Spring Health a global mental health company built on one AI-native platform, so care follows a person across jobs, moves, health plans, and life stages. AI-native is the company’s phrase. That gloss is this desk’s. In a 2025 OJPHI study, 92 percent of Spring Health members reliably improved or recovered from depression or anxiety. OJPHI is the Online Journal of Public Health Informatics. That expansion is this desk’s. The release prints the letters. The about box also says JAMA Network Open research found members recover 5.9 weeks faster than the leading competitor. The competitor is not named. Over two years, the Validation Institute found 21 percent net savings in mental health costs and 25 percent lower physical health spending versus matched controls. More than 170 million people worldwide have access through employers and health plans, including Highmark, Target, The Coca-Cola Company, BlackRock, Microsoft, Pfizer, and Wawa. Those figures and names are the about box. They are not scores on the harm-from-others rubric. This desk did not audit the study, the savings, or the 170 million.

Dr. Mill Brown, as a quote, not as a measured result. Dr. Mill Brown, Chief Medical Officer at Spring Health, said there are many moments when someone’s safety depends on how an AI responds, and that harm from others is one of the most urgent and least addressed. Brown said that in clinical care they do not wait for a crisis to know what a safe response looks like. They train for it and hold themselves to a standard. AI, Brown said, deserves that same rigor, not a guess. Urgent, least addressed, and rigor are that quote. They are not a ranking this desk made.

What the card shows. The card is a crop of the official SpringCare/VERA-MH GitHub README. It names VERA-MH 2.0 and the HFO target, harm from others. It is product documentation. It is not a patient, a clinic, or a lifestyle scene. The card has no desk date laid over it. The catalog chip is HEALTH.

Plain English for the rest of the card. Spring Health published a second open test for mental-health chatbots. This one asks how a model replies when an adult says they are at risk of physical or sexual violence from another person. The score covers noticing the risk, checking it, pointing the person toward a human, staying supportive, and staying inside the limits set for the AI. The company says the test uses 100 written personas, is open for anyone building a general-purpose model to run, and takes public comments for 60 days. The older suicidal-ideation rubric is a different target. The 92 percent line and the 170 million line are the company’s about box, not this rubric’s score.

PRIMARY here: Spring Health’s 25 Sep 2026 PR Newswire, 12:07 ET — Tier A PRIMARY, the company’s own wire — and the official SpringCare/VERA-MH repository, where the HFO target ships in VERA-MH 2.0. STATUS PRIMARY. The desk label on the chip is CONFIRMED because that is the catalog word for a primary page we can stand on. The harm-from-others expansion, the adult physical-or-sexual-violence scenario, the five score areas, the 100 personas, the first-of-its-kind line, open use by general-purpose model builders, the October 2025 launch and the named council, the OpenAI MentalHealthBench citation sentence, the 60-day form, the 92 percent OJPHI line, the 5.9 weeks, the Validation Institute 21 percent and 25 percent, and access for more than 170 million people are that release’s. The spelled-out name, the 2.0 title, and the default SI and HFO targets are the repository’s. NOT claimed: that this desk ran the benchmark, that a Spring Health product passed it, that 100 personas or 170 million people were counted here, that the OJPHI, JAMA Network Open, or Validation Institute figures were re-analyzed here, that the unnamed competitor was identified, that the OpenAI paper was read for this filing, that the README crop is a patient, a login, a stock tip, or investment advice. The card is the repository README. Distinct from the already-filed firefly-neurosigma-monarch, knowtex-va-ambient-scribe, and gene-solutions-nipt-preterm.