Meta Deploys LLM to Catch Ads Signposting Users to CSAM

Meta is pointing a large language model at the destinations ads lead to, not just the creative, after what the company describes as a tactic shift by actors trying to steer users toward child sexual abuse material hosted outside its platforms. In a post reported by TechCrunch, Meta called the ads "signposting," spots that look harmless but route users elsewhere, and said this is "a tactic it has recently seen bad actors use as they continue to change their methods to avoid detection."

The destination-analysis layer lets Meta "block websites or other destinations that break its rules and take action against the accounts behind them," the company said. Additional AI-driven scans are catching exploitation content "that earlier systems may have missed," and a new red-teaming AI agent "looks for weaknesses that bad actors could use to get around the company's protections." Meta said it has also improved detection of users who return with new accounts after removal.

The product update came with a transparency pass. Meta said it took action on 33.2 million pieces of child sexual exploitation content on Facebook and Instagram in the first half of 2026, with more than 97% detected before any user report. In India, the figure was 5.3 million pieces over the same window, with over 98% caught proactively.

The engineering push lands roughly two months after Meta's August 2026 agreement to pay up to $18 billion to settle a child safety lawsuit involving 29 U.S. states. It is also the latest in a run of child-safety alerts on our radar this week, alongside OpenAI's first teen ChatGPT usage report and Common Sense Media rating ChatGPT for Teens an "Unacceptable Risk".