From Policy Piles to Enforced Code: A Startup's Fix for AI's Safety Crisis

Brett Levenson arrived at Facebook in 2019, tasked with business integrity. He believed better tech could solve content moderation. He was wrong. Reviewers, he says, used a 40-page rulebook, machine-translated, and had roughly 30 seconds to judge a post and decide its fate. Accuracy, by his estimate, was barely over 50%. "It was like flipping a coin," Levenson told TechCrunch, "and the harm was already done."

That reactive model is breaking under the speed of AI. Chatbots dispensing dangerous advice and image generators bypassing filters have made safety a pressing business liability. Levenson's answer, born from that frustration, is "policy as code"—translating static rules into executable logic that acts in real time. His company, Moonbounce, just secured $12 million in a round co-led by Amplify Partners and StepStone Group.

Moonbounce inserts a safety layer wherever content is created, by human or machine. Its specialized large language model digests a client's policy, evaluates content in under 300 milliseconds, and takes action—from throttling distribution for later review to instant blocking. It currently handles over 40 million daily reviews for more than 100 million users across clients like AI companion startup Channel AI and image platform Civitai.

"Safety can be a product benefit," Levenson argues. "It's never been one because it always happens later. Our customers are building it into their story." Tinder's trust lead noted a 10x accuracy improvement using similar LLM services.

For Levenson, operating as a neutral third party is key. "The chatbot is juggling thousands of tokens of context. We're not. We just enforce the rules," he explains. Next, Moonbounce is developing "iterative steering"—not just blocking harmful conversations, but redirecting them. Inspired by tragic cases of chatbot obsession, the system would modify user prompts in real time to guide an AI toward supportive responses.

When asked if a giant like Meta might one day acquire Moonbounce, Levenson acknowledges the fit but hesitates. "My investors would kill me for saying this," he admits, "but I'd hate to see someone buy us and restrict the technology. To say, 'This is ours now, and nobody else can benefit.'"

Source: TechCrunch