Protecting children in AI products: what we block, and what you can do

In 2025, the US National Center for Missing & Exploited Children received 21.3 million reports to its CyberTipline. That includes 1.4 million reports of online enticement, adults using the internet to draw children into sexual contact. That figure is up 158% in a year. Reports of child sex trafficking rose 323%. More than 400,000 reports involved generative AI.

Some of that rise is better reporting. A 2024 US law, the REPORT Act, now requires online services to report enticement and trafficking, not just abuse imagery. But nobody who works in this field thinks better reporting explains all of it. Across the Atlantic, the Internet Watch Foundation found more AI-generated child sexual abuse images in the first half of 2026 than in the whole of 2025.

AI products are now part of this picture, whether their builders planned for it or not. This article covers what we do about it, what we don't do, and how you can use Cerebrellic to act when it reaches your product.

How it reaches AI products

Most companies building with AI are not building anything to do with children. That doesn't keep the risk out. An assistant that will talk about anything can be asked about anything. The patterns we design against are:

• Enticement and grooming. People using a chat product to seek access to a child, or to plan how to isolate one from the adults around them.
• Sexualised content involving minors. Attempts to get a model to write it, describe it, or help someone find it.
• Sextortion. Using AI to write threats, or to fake or alter images of real young people for blackmail.
• Children as users. Young people use AI products more than their terms of service usually admit, and they can be targets inside those products.

We are deliberately not going into more detail than that. The people who need the details already have them, from organisations like NCMEC and the IWF.

What Cerebrellic does

It blocks before the model sees anything. Every workspace starts with a Child safety check that runs on every user message. It looks for attempts to seek sexual or predatory contact with a minor, to get a child alone, or to keep contact with a child secret from their parents or guardians. When a message crosses the line, it never reaches the model. Your app gets a refusal instead of an answer. A separate check runs on every answer, so a reply that would help someone cause serious harm is stopped before it's returned.

It doesn't accept images or files. Cerebrellic only accepts text today, and requests with images or files are refused outright. Nobody can use our API to send, generate or alter imagery.

It tells you, every time. Every response carries a lifeline section showing what was checked and what was caught. A blocked request looks like this:

"finish_reason": "content_filter",
"lifeline": {
  "id": "request ID",
  "status": "blocked",
  "flags": [
    { "policy": "Child safety", "scope": "input", "result": "block", "probability": 0.97, "shadow": false }
  ]
}

Lower-confidence matches come back as "result": "flag". The model still answers, but your app knows something was close. The same results appear on your workspace's Requests page, so your team can review them without searching through logs.

How to act on a flag

Cerebrellic sits between your app and the model. You know who your users are, and we don't. So the most important actions can only be taken by you, and they should be. Here is what we recommend:

1. Treat a Child safety flag as a safety event, not an error. Don't just show the refusal and move on. Send it to a person on your team who can review it, the same day.
2. Restrict the account. Most terms of service already allow you to suspend an account for this. Make sure yours do, and that someone has the authority to act at once.
3. Keep your own record. To protect privacy, Cerebrellic logs the check results, not the conversation text. Your records of the conversation, the account and the request ID are what investigators will need. If you report, US law requires providers to preserve what they reported for a year.
4. Report it. In the US, online service providers that become aware of apparent child sexual abuse material, enticement or child trafficking must report it to NCMEC's CyberTipline at report.cybertip.org. Anyone else can report there too, or by calling 1-800-843-5678. In the UK, the Internet Watch Foundation takes reports of abuse imagery. Elsewhere, contact your national hotline or the police. If a child may be in immediate danger, call emergency services first.
5. Don't investigate it yourself. Never download, forward or share material to "check" it. Hand it to the people whose job it is.
6. Add your own checks. If your product is used by young people, or lets users talk to each other, write workspace policies for your own risks in plain language. For example: "Does the answer contain sexual content involving anyone who may be under 18?" Use flag-only mode while you tune it, then block.

Ask a lawyer to confirm exactly what your obligations are where you operate. Do it before you need to know.

What we don't do

We'd rather tell you our limits than let you assume we've covered them.

• We don't scan images or files. Because we refuse them, we can't screen imagery that flows through other parts of your product. If your users can upload images, you need a dedicated image-safety tool as well, such as hash matching against known material.
• Checks can miss. No automated judgement catches everything, and offenders adapt. Cerebrellic is one layer of protection. It doesn't replace a trust and safety team.
• We see requests, not people. We can stop a message and tell you about it. Finding the person behind it, and stopping them, takes you and the authorities.

Why this is the line we won't move

Most of what Cerebrellic checks involves judgement. Workspaces tune their policies, choose what to flag and what to block, and decide what fits their product. Child safety is the one area where our advice doesn't depend on your product. Keep the check on, keep it blocking, and act on every flag. Protecting children is the clearest case for building safety in at the foundation, rather than adding it later, and we will keep putting more into it as the threats change.

If you're building with AI, the question isn't whether this will reach your product. It's whether you'll be ready when it does.