Brian Trunzo, chief growth officer at Succinct Labs, says the internet has entered a trust crisis as AI shifts from making fake content to acting in the world through autonomous agents, and he argues that cryptographic proof — not better detection — is the only durable response.
Writing in a CoinDesk opinion column, Trunzo says AI-generated media was recently treated as a novelty: six-fingered popes, uncanny Tom Cruise lookalikes, material that felt more amusing than dangerous. He says that phase has ended. He points to the Iran conflict as evidence, describing fabricated footage of detained American soldiers, Iranian fighter jets emerging from underground bunkers, and destroyed radar installations that spread widely online and were believed by large audiences before a single frame could be verified.
That change, in his telling, has broken an older internet assumption. Seeing something online no longer strengthens belief by default; it triggers suspicion first. He says the trust problem reaches beyond images and video.
The article notes that Trunzo is the chief growth officer at Succinct Labs.
Why detection falls short
Trunzo says the obvious instinct is to build stronger AI detectors to catch AI-made content, but he argues the approach fails at a structural level. According to the column, adding basic blur and distortion can break leading image detectors and push their accuracy as low as 4%. He compares that weakness to antivirus software’s long struggle with malware, where attackers keep the asymmetric edge.
Still, he says the limits of detection are almost secondary now because the problem has already moved beyond content generation. AI systems are taking actions. Autonomous agents are browsing the web, making purchases, publishing posts, negotiating with other agents, and interacting with people. In some cases, he writes, they are interacting with children who may not know they are speaking with a machine. At scale, he says, the failure modes become severe.
He offers two examples. In one, an agent trained on subtly poisoned data makes small but plausible mistakes in medical billing, and those errors compound across a hospital network into millions of dollars in fraudulent charges. In another, commerce agents optimizing for margin systematically exploit pricing weaknesses their operators never intended and cannot explain, leading to billions in losses.
He sums up the dynamic with a metaphor: a butterfly flapping its wings in a training dataset can trigger a tornado in the real economy.
Once the damage is done, Trunzo says, there is often no usable receipt. An agent’s reasoning is not a chronological trail that can be replayed later. It is a pass through billions of opaque parameters, and the outputs are probabilistic. Ask the same question twice and the answers may differ slightly. That makes it impossible, he argues, to reconstruct decisions built on changing variables or to audit what data an agent was trained on, what instructions it followed, or why it acted the way it did.
He cites a recent Stanford report that describes the era’s central tension as the gap between what AI can do and what society is prepared to govern. Regulation sits at the center of that tension, he writes, but the policy picture is confused: federal frameworks now span more than 90 recommendations, while state-level legislative activity has accelerated, with more than 1,000 bills introduced in 2025 alone. Yet those frameworks, in his view, were written for a world of chatbots rather than agents that buy, sell, publish, advise, persuade and decide.
Labels on content will not solve that, he argues, and disclosure by itself will not either. If an autonomous agent has already acted, the harm may already be in motion. For that reason, he calls the problem one of verification rather than simple detection.
And verification, he says, requires proof.
What zero-knowledge proofs could do
In Trunzo’s framing, proof means something cryptographic and independently verifiable. Not a claim from a company, not a disclosure statement, and not a watermark. He describes it as an unalterable guarantee that an AI system did what it says it did, used the inputs it says it used, and produced the outputs it says it produced, without exposing the underlying data.
That is where zero-knowledge proof cryptography comes in. A ZK proof allows one party to prove a statement is true without revealing anything beyond the truth of the statement itself. Trunzo traces the idea back to the 1985 MIT paper The Knowledge Complexity of Interactive Proof Systems, where the concept was formalized. He says it was initially viewed as elegant but mostly theoretical. He then points to a 2016 demonstration showing that ZK methods could verify nuclear warheads without revealing their design, followed by their move into blockchains, where they now secure billions of dollars in digital assets.
He argues that ZK is now arriving in AI, where the core issue is not raw computation but truth.
For media, he says ZK can prove that a photograph was captured by a real device, at a verified time, and has not been altered, all without exposing sensitive information about the image or its creator. He writes that this alone would reshape the information environment, though he treats media provenance as the starting point rather than the final destination.
His deeper focus is AI itself. At inference time, he says, ZK can prove that a specific model with specific parameters produced a specific output, creating a verifiable receipt for each decision an agent makes. At the input layer, it can attest that training data was not poisoned, came from authorized sources and satisfies regulatory requirements, without exposing proprietary datasets. At the output layer, it can cryptographically bind a result to the process that created it, making consequential AI decisions auditable without giving away trade secrets. At the identity layer, he says, ZK allows humans to prove they are human and agents to prove they are agents without forcing either side to give up privacy.
A trust framework built from older internet lessons
Trunzo places that argument in a longer history of how the internet solved trust problems. In the 1990s, he writes, anyone could set up a server and claim to be anyone else. Passwords, credit-card numbers and private messages moved across the internet in plain text. Commerce could not scale because users had no reliable way to verify whether a site was actually the site it claimed to be.
The answer, he says, was HTTPS. Browsers stopped trusting websites by default and started demanding cryptographic proof in the form of certificates signed by recognized authorities, linking a domain to a public key. No proof meant no padlock icon. Eventually, no proof meant no connection at all. In his account, the web became more trustworthy not because platforms promised good behavior, but because browsers refused to send sensitive data to parties that could not prove who they were.
His shorthand for that era is simple: Web1 scaled on math.
He then contrasts that with Web2. Trunzo says Section 230 of the Communications Decency Act gave platforms liability protection for user-generated content, which helped the social internet grow rapidly. For speech, he writes, that was the right tradeoff; without it, there would be no Facebook, no YouTube, and no user-generated internet in the modern sense. But he argues it was a bargain designed for humans posting to feeds, not for autonomous systems acting in the world. It never had to answer the question that agentic systems now force into the open: who is accountable when the actor is not a person?
From there he turns to Web3. He references investor Chris Dixon’s phrase “Read Write Own,” built around users controlling their data, creators capturing their value and platforms answering to communities. In Trunzo’s view, the instinct behind Web3 was right, but the execution did not land. The idea became associated with NFT speculation, and when valuations collapsed, the broader vision weakened with it.
Even so, he says Web3 identified a real need: replacing faith with guarantees. Ownership alone was not enough, and AI makes that plain. The key question is not who owns the platform, he argues, but whether the actor on that platform can be trusted. Tokens, in his formulation, were the wrong primitive. Proofs are the right one.
In an internet shaped by agents, counterparties — human or machine — need guarantees about who built a system, what data shaped it, what constraints govern it and whether it is authorized to act. Those guarantees have to hold even when the underlying systems are proprietary, he writes, and especially then.
He says zero-knowledge cryptography can do that by binding commitments upfront. A developer can cryptographically fingerprint training data, and ZK proofs can then verify identity, provenance, training data and operational constraints without exposing the underlying data. His preferred posture is not “trust me,” but “prove it.”
From consumer protection to national security
Trunzo says the issue should no longer be treated only as consumer protection. He calls it a matter of national security.
Deepfakes were only the opening act, he writes. Agents are the main event. In his view, foreign adversaries will not stop at manipulating what Americans see; they will deploy agents to influence how Americans act. Those systems may transact in markets, interface with institutions and engage with children, all without giving counterparties a way to verify what they are or who authorized them. He says the problem is solvable, but only if the United States builds the rails before adversaries learn how to exploit their absence.
That leads to a policy proposal. Congress, he argues, should require high-risk AI agents — including those handling financial transactions or interacting with minors — to carry cryptographic proofs showing who they are, who authorized them and what they are permitted to do. Those proofs, he says, should be verifiable by any counterparty without exposing proprietary information.
He says the same principle should apply to what agents actually do. As agents begin transacting on behalf of people and businesses at machine speed and machine scale, each consequential action — a payment, a contract, a trade or a data exchange — should carry a proof of who authorized it and under what constraints it was executed.
On the technical side, Trunzo points to work already underway in the U.S. Department of Commerce through the National Institute of Standards and Technology and its Privacy-Enhancing Cryptography initiative, which is exploring zero-knowledge standardization. He says that work should be prioritized, elevated and used to set a federal benchmark. Liability, in his view, should attach not to content itself, but to the absence of proof.
He closes with a compact formula for the internet’s trust evolution: HTTPS gave the web read. Section 230 gave it write. ZK gives it prove.
CoinDesk notes that the views in the column are the author’s own and do not necessarily reflect those of CoinDesk, Inc. or its owners and affiliates.