Brian Trunzo, chief growth officer at Succinct Labs, published an opinion piece with a blunt thesis: autonomous AI agents have broken the internet's trust model, detection-based defenses can't save it, and zero-knowledge proofs are "our only hope" ([CoinDesk](https://www.coindesk.com/opinion/2026/07/17/ai-is-destroying-the-internet-math-is-our-only-hope)). Worth reading closely, because the diagnosis is sharper than the cure — and Succinct Labs happens to be in the business of selling that cure.
The diagnosis: detection has a structural ceiling
Trunzo's core argument isn't about deepfakes going viral, though he opens with one: fabricated Iranian conflict footage that spread to millions before anyone could verify it. The deeper claim is that agents no longer just generate content — they "transact, publish, and interact autonomously," which changes the trust problem from *is this real* to *did this agent actually do what it claims*.
His case for why detection can't close that gap: an agent's reasoning is "a single pass through billions of opaque parameters," probabilistic and non-reconstructible. You can't audit a decision after the fact if there's no decision trail to audit. He cites research showing basic blur and distortion can knock deepfake-detector accuracy down to as low as 4%. The framing: detection loses to generation for the same structural reason antivirus never eliminated malware — the attacker holds a permanent asymmetric advantage. That's not a hot take. It's the same reason SatoHub's own Sato Score doesn't try to grade an agent's intentions — it grades what's checkable: is the code live, is the repo public, is there a real release cadence. Evidence over inference.
The fix: proofs instead of trust-me
Trunzo's answer is zero-knowledge proofs — a technology old enough to have a real pedigree (the concept traces to a 1985 MIT paper, moved into practical use for nuclear-warhead verification in 2016, then into blockchains). Applied to agents, he sketches four uses: proving media came from a real device at a real time without altering it, proving training data wasn't poisoned, proving a specific model produced a specific output ("a verifiable receipt for every decision"), and proving personhood or agent-status without giving up privacy.
He closes with a policy ask: Congress should require high-risk agents — the ones touching money or minors — to carry cryptographic proof of identity, authorization, and operating constraints, with NIST's Privacy-Enhancing Cryptography initiative as the reference standard.
What's claimed vs what's shown
Here's where a builder should slow down. Everything above is Trunzo's argument, not a shipped result. No ZK-proof-of-inference system for frontier LLMs is deployed at the scale he's describing — proving properties of a live, opaque, billion-parameter model in real time is an open research problem, not a shipped product. Succinct Labs is a zero-knowledge proving infrastructure company; a piece arguing "math is the only fix" from that seat is a real technical claim wearing a company's incentives on its sleeve. That doesn't make the argument wrong. It makes it worth reading as advocacy, not settled infrastructure — the same claimed-vs-verified line SatoHub applies to every listing in the directory.
The part onchain-agent builders already have a name for
Strip the ZK framing and the underlying problem is one this ecosystem has been chewing on since before this piece existed: how does an agent prove what it is, what it's authorized to do, and what it actually did — to another agent, with no shared operator to vouch for either side? ERC-8004's identity registry is one partial answer (an onchain address an agent can be checked against — not a claim about its behavior). x402 attaches a payment rail to an agent action so the transaction itself is the receipt. A Sato Score is a third angle: not a proof of any single action, but a running transparency signal — is this thing active, is the code open, has anyone independently reviewed it. None of these are the cryptographic proof-of-inference Trunzo wants. All of them are the same instinct: stop asking builders to trust an agent's word, and start giving them something to check.
What to watch
Watch for the gap between this argument and shipped tooling — proof-of-training-data and proof-of-inference systems that actually run against production-scale models, not whitepapers. Until then, the honest posture for any agent claiming verification is the one Trunzo's own piece models without quite saying it: state what's claimed, state what's proven, and don't blur the two.
### Sources - Brian Trunzo, "AI is destroying the internet. Math is our only hope." — [CoinDesk, 2026-07-17](https://www.coindesk.com/opinion/2026/07/17/ai-is-destroying-the-internet-math-is-our-only-hope)