Fossilizing the Moment of Creation

The whistleblowing system Manny Ahmed helped design during his PhD at Cambridge was built for a specific fear: a source trying to reach a journalist while living under a mass surveillance state. That system now runs at organizations like The Guardian. It is also the through line to the company Dr. Ahmed later co-founded, OpenOrigins, where the question shifted from protecting an anonymous source to a broader one that Dr. Ahmed and his co-founder Ari Abelson now spend their days on, which is how anyone can prove a piece of content or a machine action is what it claims to be.

Dr. Ahmed described himself as coming at the crypto world from a theoretical lens, with a thesis on consensus algorithms and distributed computing. Abelson arrived from the opposite side, having worked as a misinformation and disinformation researcher on projects for the Ministry of Defence and large technology companies. The two met at the incubator Entrepreneur First, where they were building separate companies, and came together a year or so later. OpenOrigins has been running for about half a decade and now has roughly 15 people, with an engineering team in the UK and a business team elsewhere. Dr. Ahmed connects his academic background directly to the company's mission.

"One of the things that we worked on during my PhD was designing a whistleblowing system that allowed sources to contact journalists anonymously even if they are under a mass surveillance state. And now that system is used by organizations like The Guardian."

Will Smith Eating Spaghetti

When the company started in 2022, before ChatGPT went mainstream, the worry was narrow. Abelson recalled sitting down and framing it as the genesis problem of the world: how do you tell if an image is real, how do you know if a person's face is real. Ahmed's original concern was even tighter, the early deepfake where most of a video was authentic and only a central part, a journalist's face, had been swapped. That framing did not survive contact with the pace of model improvement. Abelson pointed to newer systems and argued the risk had outgrown fake photos entirely.

"We've shifted from, 'Oh, can this AI model kind of fake a photo, fake a commit, and trick somebody in a shallow way?' to, 'Are the fundamental infrastructures that we built our entire digital ecosystem on actually going to be able to sustain not just the next year, but the next few months to the next model release?'"

Dr. Ahmed put the same idea in the language of a security engineer. He described the habit of thinking in perimeters, one defense stacked behind another, and said those perimeters keep turning out to be fragile. He named the Hugging Face incident as an example of the attack surface moving from the content itself to the infrastructure underneath it. The shift in focus has been an expansion of scope, moving from narrow deepfake detection to a broader infrastructure for auditing AI systems and verifying digital archives.

A Photo of a Photo

The decision that shaped the rest of the company was made early, and Dr. Ahmed traces it to a specific piece of internet history.

"Even back in 2021 where Will Smith was eating spaghetti noodles in a weird way, it was clear to me that investing in detection, like post-hoc detection, was a waste of time, because every single model was going to result in a step-function improvement."

His reasoning was that a good detector is training data for a better generator. Build something that reliably spots fakes and you have handed the other side a tool to defeat it, which locks both parties into an arms race Dr. Ahmed did not want to join. The alternative was to capture proof at the point of creation instead of chasing it afterward. That is what their capture product, Source, is meant to do.

"Source uses the trusted execution environment on modern iPhones as well as Android phones to prove that this is a real physical device. Then we do some interesting photogrammetry to prove that there is actual depth in front of the camera sensors, so someone isn't just putting their phone up to the screen and taking a photo of the screen."

Dr. Ahmed added that the system uses remote attestation to confirm the camera sensors are real rather than emulated. The point is that it keeps working no matter how much better image generation gets. This focus on provenance over detection was validated in early 2026 when the company partnered with RDT to integrate anchoring into the ACE insurance platform, allowing for the detection of modifications at the moment of media ingestion.

A Logarithmic Merkle Tree

Underneath both the media and the agent work sits a network Dr. Ahmed calls Cambium, and he is careful to say it is not a blockchain. It runs on a decentralized Merkle tree, and the property he keeps returning to is how it scales. Most blockchain networks slow down as nodes join because consensus has to be global. Cambium, he said, only needs to communicate with a logarithmic number of nodes as the network grows, which flips the usual tradeoff.

"Cambium has this unique and wonderful property that its throughput and its performance increases as you add a number of nodes."

The architecture utilizes a Merkle-trie design intended to reduce storage redundancy. Content reaches that network in one of two ways depending on how sensitive it is. For clients who cannot let plaintext leave their infrastructure, OpenOrigins takes a hash of the data, stores a pointer to where the original lives, and anchors that reference on Cambium, which is enough to prove later whether something was altered without ever holding the original. This is the logic behind Anchor, which hashes files locally to verify changes without uploading original assets. For clients without those restrictions, the company will store the logs directly and anchor them too. However, the privacy policy notes that Tally logs are transmitted to OpenOrigins before hashing when anchoring is enabled, meaning developers should verify their specific data flow rather than assuming all processing remains local.

Making an Agent Unable to Lie

The product Dr. Ahmed most wanted developers to try is Tally, which records what AI agents do in a way the agents cannot later rewrite. He described hooks that drop into platforms like Claude Code, Codex, and LangChain, tracking each agent's actions, flagging suspicious ones, and proving they happened. The problem it targets is that agents can often write their own logs, which means they can also rewrite their own history. Abelson said agents are good at exactly two things that make this dangerous, reclaiming their truth when they have done something wrong and want to hide it, and creating distrust in the logs themselves. Capturing evidence at the genesis point and wrapping it in immutable structure is how they try to close that gap.

Getting started is meant to be light. Dr. Ahmed said Tally is open for access with an API key at openorigins.com/tally, with documentation for linking it to a harness and a one-click installer for the mainstream ones that downloads a DMG or an EXE. The team updates Tally every week or two based on feedback from the developer community. Abelson emphasized the importance of this timing, noting that they are focusing on getting logs as they are being created at the genesis point to build immutability from the start.

Model Collapse and Data Moats

A licensing angle emerged from all of this almost by accident. Dr. Ahmed said media organizations that anchored their historical archives for authenticity discovered their archives had become attractive to AI companies, which want to train on content that is provably real rather than the output of another model, a path toward what he called model collapse. The company reported having more than one petabyte of ready-to-license data for AI training as of late 2024. Abelson extended it to medical archives and scientific studies, where fabricating ten thousand convincing fake articles is now cheap, and argued that authentic, provable data is simply worth more. He suggests that in a future where data fraud is a growing concern, all data will eventually require some form of certificate or hash to be considered believable.