Three incidents. Zero independent witnesses. That is the entire story, and the market is only beginning to feel it.
Over the past seventy-two hours, a report from Crypto Briefing has been moving through the inboxes of every risk manager who holds a position adjacent to artificial intelligence. The report names OpenAI, Anthropic, and Meta. It says that a series of incidents reveals a dangerous gap in AI oversight. It does not give timestamps. It does not give logs. It does not give transaction hashes. It does not even say what the incidents were. On a first pass, that absence looks like weak journalism. On a second pass, it looks like the exact symptom the report is trying to describe.
I have worked in this intersection for eighteen years. I have watched unverifiable narratives move markets more violently than verifiable data ever did. In 2026, my team and I built a signal-monitoring system to separate human trading from machine trading. We found something we called Algorithmic Herding: clusters of autonomous agents executing in near-perfect synchrony, often within the same second, without a single human command. We measured that more than thirty percent of daily volatility in the crypto assets we tracked was tied to non-human actors. The terrifying part was not the volume. It was the opacity. When those agents moved, there was no public record of why. There was no independent system watching them. There was only the market’s collective panic after the fact.
This is why the OpenAI, Anthropic, and Meta story should not be filed under AI news. It belongs under market-structure news. The market is finally staring at a gap that has been there since the first frontier model was deployed: no independent witness, no audit trail, no way to prove that a model did what the lab claims it did. The market did not crash. It woke up.
Let me establish something important. OpenAI, Anthropic, and Meta are not just AI companies. They are the closest thing this industry has to digital sovereign utilities. Their models are the rails on which a growing share of enterprise decisions, code generation, and financial automation runs. When one of these labs updates a model, the change can ripple through every downstream product within minutes. A small change in the prompt policy of a frontier model can alter the output of thousands of AI agents that are executing trades, writing risk reports, or managing token allocations. There is no circuit breaker for that cascade.
The current oversight model is entirely self-referential. Each lab has a safety team, and each safety team reports to the leadership of the lab. That is not independent oversight; that is internal risk management. It is useful, but it has all the same incentive problems that every self-audited protocol in crypto had before its exploitations. When a team is rewarded for shipping, safety findings are treated as blockers. When a team is funded by the company’s valuation, disclosure of a serious incident becomes a valuation problem. The result is a structural bias toward silence. That bias is not evil. It is structural. It is the same reason that a liquidity mining program in DeFi is more about subsidizing the TVL number than about finding real users. In both cases, the measurable incentive dwarfs the unmeasurable truth.
I also need to address the credibility question. Crypto Briefing is an industry publication, not an academic journal. You can question its editorial agenda. But the report’s lack of detail should be treated as evidence, not as a reason to ignore it. In a healthy oversighting regime, events would be independently documented at the moment they happen. In the current regime, the only documentation is whichever employee feels comfortable leaking it to whichever reporter asks first. The absence of a complete record is exactly what a complete record would reveal.
Now the core of the problem. Call it oversight latency. By that I mean the time between an event and the first independent verification of that event. In high-frequency trading, latency is measured in microseconds; the lower it is, the more alpha you capture. In crypto, latency is also tiny: a transaction is public the moment it enters the mempool, and an independent observer can verify almost any claim with a block explorer. In AI, oversight latency is not measured in seconds. It is measured in the time it takes a corporate spokesperson to say that they cannot comment on an ongoing investigation. That latency is effectively infinite, because the lab is the only party with access to the model’s internals. Infinite oversight latency is catastrophic when the models themselves are market participants.
Consider a real scenario from my own monitoring. In early 2026, my team detected a strange wave of volume across several small-cap liquid pools. There was no catalyst in the news feed, no whale footprint, no on-chain anomaly in the traditional sense. What we saw was a pattern: dozens of AI trading agents began selling the same token within the same seven-second window. They did not know each other. They were different models from different vendors. But they had all been updated with the same training tactic, a preference-learning adjustment that made them interpret a certain type of on-chain liquidity drop as a signal to exit. The update had been rolled out quietly by a lab that we later identified. The agents did not collude; they were synchronized by identical training data and identical inference logic. That is Algorithmic Herding. It produced a flash crash in a token that had no reason to fall.
Now ask the question: if that lab had suffered an incident that day, who would have been able to prove the connection? The agents’ logs belonged to their operators. The model’s update notes belonged to the lab. The exchange data was public, but nobody could map the causal link without deep cooperation. The market was left to explain the crash with rumors. That is the oversight gap. It is not a philosophical problem. It is a missing audit trail between model updates and market outcomes.
Let’s make this concrete. What would a real independent oversight layer for OpenAI, Anthropic, and Meta look like? I am not talking about a government commission. I am talking about cryptographic infrastructure. There are three layers that need to exist.
The first layer is a model registry. Every deployment of a frontier model should be accompanied by a public commitment: a hash of the weights, a hash of the tokenizer, a hash of the base configuration, and a signed statement describing what the lab claims the model can and cannot do. This is not code disclosure. It is provenance. The lab keeps the weights private. The world gets a verifiable fingerprint. Later, when an auditor wants to verify a claim about the model, they can reconstruct the model under a controlled protocol and compare the fingerprint. If the hash matches, the model is what the lab says it is. If the hash does not match, then the model was silently changed, and now there is proof.
The second layer is signed inference receipts. Every time a model is used to make a consequential decision, a trade, a credit decision, a medical recommendation, a security clearance, the inference should produce a cryptographic receipt. The receipt includes the input hash, output hash, model version, and a short policy trace that explains the chain of reasoning. The receipt is published to a public ledger or a permissioned witness set. No weights are revealed. No private inputs are revealed. The proof only tells the world that a particular model version produced a particular output at a particular time. This is basic hash-chaining, the same machinery that powers blockchain explorers.
The third layer is an aggregated incident feed. When a lab confirms an anomaly internally, it should be legally and technically required to publish a signed notice. The notice should contain the model version, the affected use case, the estimated reach, and any associated on-chain transactions. The market can then instantly estimate severity. The lag between the incident and the notice becomes a public metric, call it the incident latency. Once that metric is public, it will be priced into everything. Labs with fast disclosure will earn a governance premium. Labs with slow disclosure will suffer a governance discount. That is the only kind of regulation that scales in a global market.
Back in 2020, I ran a liquidation bot on Compound Finance. I spotted a flaw in the health factor calculation during a flash-loan attack and captured over one hundred thousand dollars in fees while others got liquidated. That was not special. It was simply a case where a public protocol could be audited by anyone with enough patience. The code was on-chain. The state was on-chain. The failed transactions were on-chain. The exploit was visible before anyone announced it. In AI, none of that exists. There is no Chainlink feed for model behavior. There is no block explorer for inference logs. There is no mempool for training data. The entire system is one giant off-chain black box that is now being asked to manage an increasing slice of the financial system.
Some will say the answer is regulation. But the current regulatory apparatus is not built for this. Regulators are trying to inspect frontier models in the same way they inspect banks: by sending examiners to read documents. That will fail. By the time an examiner understands a model version, the lab will have replaced it. The speed differential is too large. In crypto, we solved this by moving the trust anchor from institutions to mathematics. The same shift is required for AI oversight. You cannot inspect a border collie with a spreadsheet; you need a motion detector. For AI, the motion detector is a cryptographic audit trail.
Now the contrarian turn. The obvious recommendation from the Crypto Briefing report is that we need independent AI oversight institutions. I disagree. I think a centralised independent oversight institution would create exactly the kind of systemic risk the report fears.
Follow the logic. A new agency with jurisdiction over OpenAI, Anthropic, and Meta would need access to model weights, training sets, inference logs, and internal deployment graphs. That is the most sensitive industrial data on planet Earth. It would become an intelligence target for every state actor, every criminal cartel, and every rogue lab. One compromise inside that agency would give an attacker access to all of it. You would not be reducing centralisation; you would be constructing a single boring point of failure. In blockchain terms, it is a trusted setup with one key. Every security engineer knows what happens to trusted setups over time. The key leaks.
The solution is not more oversight power. The solution is less oversight power delivered through more oversight infrastructure. Publish signed receipts. Publish model fingerprints. Publish incident notices. Then let every independent researcher, journalist, and market participant be a witness. That is how you move from an accountability gap to an observability gap. The collective panic in the market right now is not because OpenAI, Anthropic, and Meta made mistakes. Every technology company makes mistakes. The panic is because nobody can verify the size, the timing, or the consequences of those mistakes. And that is a structural condition that no new committee can fix.
Here is what I am watching next. The first major AI lab to deploy a public model registry with signed inference receipts will separate itself from the herd. The first institutional buyer to require machine-readable incident logs from its AI vendors will set the procurement standard for the next decade. If no one moves first, the market will perform the audit itself. When the next AI-driven flash crash hits, the on-chain trace will be the only witness. The question is whether the labs build that witness before the first systemic loss.
Three incidents. Zero independent witnesses. In a transparent system, that sentence would be impossible. The market is not asking for another regulator. It is asking for a block explorer for artificial intelligence. The question is not whether that explorer gets built. It is whether OpenAI, Anthropic, and Meta decide to build it themselves, or whether they are forced to by the trail of their own next incident. I would not wait. The market stopped waiting the moment it saw the latency.
