Meta deploys AI to find ads linking to child abuse

Meta is deploying a new large language model system designed to combat 'signposting,' a tactic where bad actors use seemingly benign advertisements to redirect users to external websites hosting illegal child sexual abuse material. By analyzing the ultimate destinations of these ad links rather than just the surface-level ad content, the system allows Meta to block malicious external domains and terminate the accounts responsible for them.

Alongside the LLM, the company is implementing a red-teaming AI agent to proactively probe its own defenses for vulnerabilities before bad actors can exploit them. Meta is also utilizing supplementary AI scans to catch previously missed content and upgrading its detection of ban-evaders who attempt to set up new accounts. These tools build on Meta's existing automated moderation, which flagged 97 percent of the 33.2 million pieces of child exploitation content the company addressed on Facebook and Instagram in the first half of 2026. In India, proactive detection reached 98 percent of the 5.3 million pieces of content acted upon during the same timeframe.

The rollout follows intense regulatory pressure, including an August agreement where Meta settled a child safety lawsuit with 29 U.S. states for up to $18 billion. For trust and safety practitioners, Meta's shift toward destination-based analysis and automated red-teaming represents a significant evolution in content moderation. Instead of relying solely on static media classifiers, engineers must now deploy dynamic LLMs capable of evaluating contextual intent and external web environments. This multi-layered approach highlights how modern platform defense requires continuous, automated vulnerability testing to keep pace with adversarial adaptation.