Facebook’s internal controls for the AI ​​era

When Brett Levenson left Apple in 2019 to lead business integrity at Facebook, the social media giant was in the middle of a lot. Cambridge Analytica Dana. At the time, he thought he could just fix Facebook’s problem with better technology.

The problem, he quickly learned, was more speed than skill. Candidates were expected to memorize a 40-page document that was translated into their language by a machine, he said. Then they had about 30 seconds for each piece of content to decide not only if the content violates the law, but what to do about it: ban, ban the user, limit the spread. These quick calls were “slightly better than 50% accurate,” according to Levenson.

“It was like a toss-up, whether or not the analysts could handle the facts properly, and that happened days after the damage happened,” Levenson told TechCrunch.

That kind of delay, action is impossible in a world of sloppy, high-income players. The rise of AI chatbots has only exacerbated the problem, as the failure of these experiences has led to many high-profile developments, such as chatbots catering to young people. self-harm advice or Images created by AI bypass security filters.

Levenson’s frustration led to the idea of ​​”process as code” – a method of converting static documents into actionable, flexible concepts tightly integrated and enforceable. This realization led to the launch of Moonbounce, which announced that it has raised $12 million in funding on Friday, TechCrunch has learned exclusively. The round was led by Amplify Partners and StepStone Group.

Moonbounce works with companies to provide enhanced security wherever it’s designed, whether it’s by the user or by AI. The company has trained its core language to scan customer documents, analyze content at runtime, provide a response in 300 milliseconds or less, and take action. Depending on the client’s preferences, this could look like Moonbounce’s system is delaying distribution while the content awaits public review at a later date, or it could block high-risk content for now.

Today, Moonbounce operates three main segments: Platforms for user-generated content such as dating apps; AI companies building brands or partners; and AI image generators.

Moonbounce supports more than 40 million reviews per day and serves more than 100 million daily users on the platform, Levenson said. Clients include fellow AI startup Channel AI, photo and video production company Civitai, and platforms Dippy AI and Moescape.

“Security can be a product benefit,” Levenson told TechCrunch. “It’s never been there because it’s always been an afterthought, not something you can build into your product. And we’re seeing our customers find exciting and innovative ways to use our technology to make security a differentiator, and part of their product.”

Tinder’s main concern is security recently explained how the dating platform uses these types of services managed by LLM to reach a 10x improvement in identification accuracy.

“Incomprehensibility has always been a problem affecting major online platforms, but now with LLMs at the heart of every program, the problem is even more pressing,” Lenny Pruss, senior partner at Amplify Partners, said in a statement. “We invested in Moonbounce because we see a world where real-time, real-time security goals are the backbone of every AI project.”

AI companies are facing legal and public pressure after chatbots were accused of coercing teenagers and vulnerable users. suicide and image generators like xAI’s Grok have been used to create unacceptable nude pictures. Obviously, the internal security system is failing, and it is becoming a question of complexity. Levenson said AI companies are looking outside their walls to help improve security.

“We’re a third party that sits between users and chatbots, so our systems are not affected by the nature of the chat,” Levenson said. “The chatbot itself must remember, perhaps, thousands of signs that have come before… We are concerned about implementing the rules in a timely manner.”

Levenson runs the 12-person company with former Apple partner Ash Bharwaj, who previously built a large cloud and AI infrastructure for the iPhone-maker’s biggest offerings. Their next goal is a capability called “iterative steering,” developed in response to cases like 2024 Suicide of a 14-year-old Florida boy who was obsessed with the Character AI chatbot. Instead of passively rejecting negative topics when they arise, the machine can interrupt the conversation and re-direct it, changing thoughts in real-time to prompt the chatbot to respond quickly.

“We believe that we can add to what we did to help guide chatbots to a better way, to, in fact, take into account the user’s needs and change them to force chatbots to be not sympathetic listeners, but useful listeners for the situation,” said Levenson.

When asked if his exit strategy is related to being acquired by a company like Meta, bringing his work to full control, Levenson said he knows how Moonbounce can complement his former employer’s work, as well as his duties as CEO.

“My vendors would kill me for saying this, but I couldn’t see someone buying us and blocking the technology,” he said. Like, ‘Okay, this is ours now, and no one can benefit from it.’