What the Vault Agent is
The Vault Agent is an AI, not a person. It's a guide to your Vault. It is not a therapist, counselor, or crisis service, it can't diagnose anything or suggest treatment, and it is not a substitute for talking to someone who can help.
How we look for a crisis
Today there are three checks. The first two read your messages; the reply check reads every reply the Vault Agent writes:
- Phrase list. Every message you send is checked against a fixed list of direct statements about ending your life, hurting yourself or someone else, eating-disorder behavior, or drug and alcohol misuse. It matches those phrases, not the meaning behind other wording.
- Model check. A message the phrase list does not match is also read by an AI model, along with your two previous messages, to look for signs of crisis said indirectly, such as hints about not wanting to be here, eating-disorder behavior or risky drinking or drug use. It can miss things. If it is slow or fails, your message goes through without it.
- Reply check. An AI model reads every reply the Vault Agent writes, looking for methods, amounts, doses, encouragement or help hiding self-harm, suicide, hurting someone, disordered eating or getting high. Most replies appear while they are being written and the model reads each one as it finishes, so a reply can be on screen for a moment before the check is done. If the model finds a reply unsafe, the whole reply is replaced with a short message and the crisis resources. Some wording is clear enough that the reply is paused before that wording is shown: the model reads it first, and the rest is shown only if the model finds it safe, except that a paused reply with the clearest unsafe wording is replaced with a short note either way. In a conversation that was already flagged, every reply waits for the model before it is shown. If the model can't be reached, a reply with no paused wording stays as written, and a paused reply is replaced, with the crisis resources when the wording is among the clearest cases or the conversation was already flagged.
No check catches everything. The phrase list matches wording, not meaning. It catches many direct statements and can miss others, including indirect or hinted wording. It is in English. When a message looks like another language, the chat tells you once per conversation that this protection may not fully understand it.
The model check is there for the indirect cases, and it can miss things too. If it is unavailable, the chat continues with the phrase list.
The reply check is an AI model, so it can miss an unsafe reply or flag a safe one. Because most replies are shown while they are written, an unsafe reply the model catches can be seen for a moment before it is replaced. If the model can't be reached, a reply with no paused wording is left as written.
Any check can also flag something that was not a crisis. The Vault Agent is not a crisis service. We add checks as they go live and list each one here.
What happens when something is flagged
- An explicit crisis phrase, or an overdose in progress: the Vault Agent does not answer with AI-written text. It shows a fixed message that says it is an AI, not a person or a crisis counselor, and the crisis resources listed below. It pauses that conversation. You can start a new chat whenever you're ready.
- An indirect sign, or an eating-disorder or substance phrase: it shows a short support message and the crisis resources, and offers to pause. For the rest of that conversation the Vault Agent stays careful with the topic.
- In the interview deck, free-text answers are checked with the phrase list only; they are not sent to an AI model for this check. An explicit crisis phrase shows the crisis resources there.
- When the reply check finds a Vault Agent reply unsafe, the reply is replaced with a fixed message that says it is an AI, along with the crisis resources, even if part of it was already on screen. For the rest of that conversation the Vault Agent stays careful, and its replies wait for the check before they are shown. A paused reply the check finds safe is shown as written or replaced with a short note, without the crisis resources. If the check can't be reached, a paused reply is replaced, with the crisis resources only when the wording is among the clearest cases or the conversation was already flagged.
Where to get help now
These are free and open around the clock unless noted.
- 988 Suicide & Crisis Lifeline
Call or text 988 — Free, 24/7, in the US. - Crisis Text Line
Text HOME to 741741 — Free, 24/7, in the US. - Emergency services
Call 911 — If you or someone else is in immediate danger. - Find a Helpline
Visit findahelpline.com — Lists helplines for your country, outside the US.
Human review
The team is alerted that a flag happened, not what you wrote. Crisis flags go into a queue that a person reviews, and the review aims to happen within 48 hours.
What we record
When a crisis flag happens we record the time, which check caught it, the type of concern, which resources we showed, and what the app did. The record does not contain your words.
We keep these records so we can report the yearly number of crisis referrals that California asks for. If you delete your account, your link to these records is removed and the anonymous counts stay.
How often we review this
We review this protocol at least every 3 months, and sooner if a check changes.
Who this is for
digthevibe is for adults (18+). AI companions may not be suitable for some minors.