Meta has started using a large language model to find advertisements that look innocent on Facebook and Instagram but direct people to child sexual abuse material elsewhere on the web, TechCrunch reported Wednesday. The system is one of several AI tools Meta announced to detect child exploitation content, test its own defenses and catch offenders who return under new accounts.

What Happened

The new model is aimed at a tactic Meta calls “signposting,” in which an ad contains nothing illegal itself but is built to steer viewers to illegal content hosted off Meta’s services. Instead of judging an ad only by its text and images, the model examines where the ad sends people. That lets Meta block the destination sites and take action against the accounts connected to them, according to TechCrunch.

Meta also said it has added AI scans designed to surface child exploitation material that its earlier detection systems missed. A third tool is an AI agent that red-teams Meta’s own safety measures, probing them for weaknesses so the company can find exploitation tactics before offenders use them at scale.

RelatedNew York City Comptroller Urges Anthropic to Delay IPO Over Safety and Governance Concerns

The company said it has also improved its ability to identify repeat offenders who open new accounts after being removed.

Why It Matters

Signposting is designed to defeat content-based moderation. Classifiers that scan an image or a caption for prohibited material can pass an ad that contains none, even when the ad exists to send people to it. Following the link to its destination moves enforcement from the ad to the network behind it, which is why Meta says it can now act on the outside sites and the linked accounts as well.

The red-teaming agent reflects a broader shift in trust and safety work, where companies use AI to attack their own systems rather than waiting for offenders to reveal new methods. Meta did not say which outside groups, if any, helped build or evaluate the tools.

The scale of the problem is large. Meta says it took action on 33.2 million pieces of child sexual exploitation content across Facebook and Instagram in the first half of 2026, and that its systems found 97% of it before anyone reported it. In India, it acted on 5.3 million pieces over the same period, 98% of them proactively.

Context

The announcement comes as Meta faces heavy legal pressure over children’s safety. In August, Meta agreed to pay up to $18 billion to settle a lawsuit brought by 29 U.S. states over social media harms to children, according to TechCrunch.

It also follows a string of youth safety features Meta introduced this year. TechCrunch listed Instagram alerts tied to self-harm content in February, preteen accounts on WhatsApp in March, parental controls for Meta AI in April and WhatsApp parental controls for teens in September.

The proactive detection figures are Meta’s own and have not been independently audited. A high proactive rate measures how much of the content Meta found was caught by its systems first; it does not show how much material went undetected.

What to Watch

The test for the signposting model is whether it reduces the number of exploitative ads that reach users, which Meta’s next transparency figures should begin to show. Regulators and plaintiffs in the remaining child safety cases against Meta will also weigh whether tools like these change the company’s exposure. Meta has not said whether it will share the signposting detection approach with other platforms.