MistralAI-Training

• Training data collection: Mistral says the crawler helps build datasets for generative AI training.
• Easy opt-out: Mistral says you can disallow this user agent in robots.txt.
• Clear separation: Separate tokens let you set different rules for training, indexing, and user visits.
• Documented identity: The user-agent links to Mistral’s crawler documentation. That helps identify the claimed operator, but it doesn’t prove a request came from Mistral.

Anyone can copy a user-agent string. A request with the MistralAI-Training token is a claim of identity, not proof of origin.

• Spoofing: An unrelated scraper can send the same string and look like Mistral in your logs.
• Wrong rules: Trusting the user-agent alone can apply the wrong policy to unrelated traffic.
• Mixed-up agents: Mistral’s three agents have different purposes. An IP list for one doesn’t verify another.
• High volume: Heavy traffic can point to collection activity, but other scrapers look the same. Volume alone isn’t proof.

A copied user-agent doesn’t give anyone access to Mistral’s systems or datasets. The impact on your site comes from the requests themselves, such as added load or noisy logs.

Mistral describes MistralAI-Training as a crawler for building training datasets. A matching user-agent in your logs doesn’t prove the request came from Mistral.

• Public content: Crawlers can reach public pages. Mistral’s documentation doesn’t define which content it collects.
• Training isn’t search: Mistral says Training doesn’t feed search indexing or live Vibe answers. Visibility in Mistral’s search depends on the other agents.
• Crawl volume: Mistral’s page gives no crawl rate or schedule. Use your server logs to see what reaches you.
• Past data: Mistral’s page doesn’t say what happens to content collected before a robots.txt change. For details, check Mistral’s terms or contact the company.

1. Check the user-agent. The documented value is Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; MistralAI-Training/1.0; +https://docs.mistral.ai/robots). Match on the MistralAI-Training token rather than the full string. Treat it as a claim, not proof.
2. Don’t borrow the other IP lists. The MistralAI-User and MistralAI-Index lists are for those agents. A match doesn’t authenticate a Training request.
3. Recheck Mistral’s documentation. The official crawler page has the current identifiers. Check it again before you act or publish.
4. Look at behavior. Review requested paths, request rates, and response patterns. Normal page fetching doesn’t prove identity.
5. Compare against robots.txt. If requests continue after you disallow the token, check your syntax, caching, deployment, and the exact token. Also check where the requests come from.
6. Use reverse DNS carefully. Mistral doesn’t document a reverse-DNS pattern for Training. A lookup can be a useful signal, but it isn’t confirmation.

It depends on your policy on AI training, content rights, and what you see in your traffic. A user-agent match isn’t a security assessment.

Consider blocking when:

• Your content is subscription-based, licensed, or commercially sensitive
• Your policy says your content shouldn’t train AI models
• Request volume affects performance or data transfer costs
• You can’t verify the traffic well enough for your policy

You can block training and still allow Mistral’s user visits and indexing. Set rules for each token separately, and check your full robots.txt for wildcard or overlapping rules. Blocking Training controls that crawler only. Mistral doesn’t say how it affects visibility elsewhere.

Use the dedicated robots.txt token. Mistral says you can disallow this user-agent.

Block the whole site:

User-agent: MistralAI-Training
Disallow: /

Block selected paths:

User-agent: MistralAI-Training
Disallow: /premium/
Disallow: /members/

These rules state your preferences. robots.txt isn’t access control, and it relies on the crawler to comply. Don’t use it alone to protect confidential content.

Other controls:

• Keep rules separate: Write agent-specific groups and review wildcard rules so the result matches your intent.
• Protect sensitive content on the server: Use authentication and authorization. Rate limiting and bot management can add another layer.
• Check your logs after changes: Confirm requests change, and look into traffic that continues.

Treat MistralAI-Training requests as unverified until you can confirm where they come from. Mistral documents it as a training-data crawler with a robots.txt opt-out, but lists no IP file for it.

Our advice: decide based on your content policy, not on security risk. Block it if you don’t want your content used for model training, and allow it only if that’s a deliberate choice. Since user-agent strings can be spoofed, don’t rely on the token alone to tell real Mistral traffic from impostors.

This is where we can help. With DataDome Bot Protect, you can see which AI training crawlers visit your site and enforce your policy with intent-based detection, even when robots.txt is ignored or a user-agent is faked. You set separate rules for training, indexing, and user-triggered Mistral traffic.

Whatever you decide, recheck Mistral’s documentation periodically, and keep sensitive paths behind authentication and authorization.