The Quiet Announcement That Should Worry Every Builder
Last week, a headline crossed my feed that most of my Telegram groups scrolled past without a second glance: Anthropic, OpenAI, and Google collaborate on AI safety standards. Three logos, one sentence, zero details. No whitepaper. No governance charter. No list of the actual standards being discussed. Just the news that the three most powerful private AI labs on Earth have agreed to agree on something — and that Crypto Briefing, of all outlets, was the one carrying the wire.
That last part is what made me stop. I have spent the better part of a decade auditing smart contracts, and I have learned that the publication carrying a story is often more revealing than the story itself. When a crypto-native outlet reports on a closed-door AI safety pact among centralized labs, something is being signaled to the Web3 audience that the mainstream financial press is not being told. My instinct as a Tech Diver is to ignore the headline and start reading the seams. There are no seams here. That is precisely the anomaly.
Because here is what three companies with a combined model-market share north of seventy percent have just done: they have positioned themselves as the arbiters of what "safe AI" means, before any regulator, any academic body, or any open-source community has been invited to the table. I have watched this exact choreography before — not in AI, but in DeFi, in Layer2, and in Bitcoin mining. Each time, the language was different. Each time, the outcome rhymed.
Context: Why Three Competitors Are Suddenly Holding Hands
Let me set the stage properly, because the surface story is misleadingly simple.
Anthropic builds Claude. OpenAI builds GPT. Google builds Gemini. On any given Tuesday, these three companies are fighting for the same enterprise contracts, the same talent, and the same benchmark bragging rights. They are not natural collaborators. They are, by design and by incentive, adversaries. So when adversaries announce a joint safety initiative, the naive reading is "the industry is maturing." The experienced reading is "a standard is being drafted, and the drafters intend to own the referee."
What is driving this? Three pressures converging at once.
First, regulatory gravity. The EU AI Act moved from proposal to enforcement reality over the past two years. China's model registration regime requires safety documentation. In the United States, the NIST AI Risk Management Framework has been quietly evolving from voluntary guidance toward something closer to de facto expectation. Every large lab now faces a compliance cost curve — and the cheapest way to shape a cost curve is to define the metrics it uses.
Second, public sentiment. The gap between what AI can do and what the public understands about AI is widening, not narrowing. Legislators who cannot explain a transformer will still be asked to regulate one. When that happens, they will reach for the only documents available: the ones the industry wrote for them.
Third, and this is the one nobody says out loud, competitive self-interest. If you are Anthropic, safety is your brand. If you are Google, safety is the dimension on which you can plausibly claim parity with a faster-moving competitor. If you are OpenAI, safety is the language you speak to enterprise buyers who are nervous about deploying language models in regulated industries. A shared safety standard is not a concession to ethics. It is a moat, poured in the shape of an ethics committee.
The Crypto Briefing angle matters here because the Web3 audience has a specific stake in this story that the general press is missing entirely. The most interesting frontier of AI right now is not a bigger model — it is an agent that can hold a wallet, sign a transaction, and operate autonomously on-chain. If the safety standards governing those agents are written by three companies whose business models depend on centralized inference, then the decentralized AI thesis gets strangled before it finishes its first sentence.
I have seen this movie. In 2017, I spent three months inside the Geth client, line by line, tracing how the GHOST protocol's fork-choice rules were actually implemented versus how the yellow paper described them. I found three edge cases in block header validation that could split a chain under high latency. The lesson I carried out of that audit was not about Ethereum specifically. It was that the gap between a stated rule and its implementation is where power actually lives — and whoever controls the implementation controls the outcome, regardless of what the stated rule says.
That is the question I want to hold onto as we look at this AI safety pact. Not "is safety good?" Of course safety is good. The question is: who implements it, and against what failure modes is it silently optimized?
Core: Reading the Seams of a Standard With No Text
The Technical Route Nobody Will Describe
When a blockchain protocol announces a new consensus mechanism, the first thing a competent auditor does is ask for the specification. Not the blog post. Not the tweet thread. The specification — the document that defines, in unambiguous language, what the system does under adversarial conditions.
This announcement has no specification. What it has, based on the limited information available, is a statement of intent. That gap is not a minor omission. In safety engineering, an intent without a spec is a liability, because it lets every party define "safe" to mean whatever suits its current deployment.
Let me be concrete about why this matters. "AI safety" is not a single technical problem. It is at least four distinct ones that frequently trade off against each other:
- Alignment — ensuring a model's behavior matches human intent across distribution shifts.
- Robustness — ensuring a model resists adversarial inputs, jailbreaks, and prompt injection.
- Capability evaluation — determining when a model crosses a threshold that requires additional oversight.
- Content governance — filtering harmful outputs at the interface layer.
A standard optimized for content governance looks nothing like a standard optimized for capability evaluation. The first is about what users see. The second is about what companies are allowed to build. If the three labs converge on the first because it is cheap and PR-friendly, and quietly deprioritize the second because it constrains their own roadmaps, then the "safety standard" is theater with a compliance sticker.
Which of these four is on the table? We do not know. And the fact that we do not know — after a public announcement — is itself the finding. Audit the intent, not just the syntax. Right now, the intent is opaque and the syntax does not exist.
The Commercial Architecture of a Safety Moat
I want to walk through the commercial logic carefully, because this is where the crypto parallel becomes almost uncomfortable.
In 2020, during DeFi Summer, I reverse-engineered Uniswap V2's core contracts to understand the slippage mechanics of the constant product formula. In the process, I found a subtle rounding error in the price oracle calculation that affected low-liquidity pairs disproportionately — retail traders got hit, whales did not. I wrote it up in Thai and English, and held a live webinar that reached over two thousand local investors.
The structural observation from that episode has stayed with me: the party that defines the measurement standard determines who bears the cost of its imprecision. Uniswap's formula was not malicious. It was simply optimized for the failure modes its designers were thinking about. Retail traders were an afterthought in the design, and the design reflected that.
Now map that logic onto AI safety standards. If the three largest labs define the safety metrics, they will — inevitably, structurally, without anyone intending it — optimize those metrics for the failure modes they are thinking about. The metrics will be calibrated against the deployment patterns of frontier commercial models. They will not be calibrated against, say, a decentralized swarm of small open-source agents transacting on-chain, because none of the three labs runs that workload.
The result is not a conspiracy. It is something more insidious: a standard that is technically neutral and structurally exclusionary. Small labs, open-source projects, and decentralized AI networks will find that compliance with the emerging standard requires infrastructure, documentation, and audit capabilities that only well-capitalized incumbents possess. They will be measured against a yardstick built for someone else's body.
This is exactly what happened with the Bitcoin ETF custody architecture in 2024. When I analyzed the multi-signature and MPC key generation processes of the major custodians, I found centralization risks that were not bugs — they were features of the institutional architecture. The custody model was designed to satisfy auditors and regulators, and satisfying auditors and regulators requires designated custodians, and designated custodians concentrate control. The decentralized ethos of Bitcoin was preserved in the marketing and quietly unbundled in the implementation.
I published a whitepaper on this — Centralization Risks in Tokenized ETFs — and the response from institutional readers was interesting. Almost none of them disagreed with my technical findings. Most of them said some version of: "Yes, but this is the only architecture that regulators will approve." Which is precisely the circular logic we should expect from AI safety standards. The standards will be shaped to be regulator-friendly. Regulator-friendliness will require institutional forms. Institutional forms will require incumbents. And the incumbents will be the three companies that wrote the standards.
The loop closes. Code is law, but trust is the currency — and in this case, the three labs are simultaneously minting the currency and holding the reserve.
The Competitive Landscape: Who Is In the Room, and Who Is Not
The most revealing thing about a closed-door meeting is the guest list. Here is what we can infer from the announcement as stated.
Anthropic's position is the most legible. Safety is its core differentiation. Every product decision Claude makes is filtered through a safety narrative. Participating in a joint standard reinforces its brand — and, more importantly, gives it a platform to argue that the safety bar should be set where its own models already clear it. There is a version of this that is genuinely principled. There is also a version that is competitive positioning in ethical drag. Anthropic is mature enough as an organization that both versions are probably true at once.
OpenAI's position is more complicated. As the market leader by deployment, it has the most to lose from any standard that constrains capability deployment. But it also has the most to gain from a standard that legitimizes large-scale deployment in regulated sectors. Its incentive is to participate enough to shape the standard, and to shape the standard enough that it constrains competitors more than itself. This is not cynicism — it is the ordinary physics of standard-setting.
Google's position is the most defensive. In the frontier model race, Google has been in the position of a company with extraordinary research depth and a recurring difficulty converting that depth into product dominance. Safety standards offer a non-capability axis on which Google can plausibly lead. If the industry converges on safety as a primary purchase criterion, Google's deep bench in alignment research becomes a commercial asset rather than a cost center.
Now: who is not in the room? The most conspicuous absence is Microsoft — OpenAI's largest investor and the operator of the Azure infrastructure that hosts a significant fraction of frontier AI workloads. Microsoft's exclusion from the named parties is either a reporting gap or a meaningful signal. If Microsoft declined to participate, that is worth understanding. If Microsoft was not invited, that tells us the collaboration is being deliberately scoped to a specific competitive configuration.
Meta is absent. xAI is absent. Every open-source consortium is absent. Every academic safety organization — the Alignment Forum at Stanford, the Center for AI Safety, the Machine Intelligence Research Institute — is absent from the named list. Every non-Western lab is absent. If the standard that emerges is intended to have global applicability, its drafting table is not global.
This is not an abstract concern. It is the same concern I raised in the Terra/Luna post-mortem. When the UST rebalancing algorithm failed, the failure was not in any individual actor's cupidity — it was in a design that had been optimized by a small team for a set of assumptions that the broader market did not share. The design was elegant. The assumptions were wrong. And the people who bore the cost had no seat at the table where the assumptions were set.
I spent six weeks in 2022 dissecting that algorithm for a Thai community that had lost real money. I wrote five posts and hosted weekly Discord AMAs. The technical content mattered, but the thing that mattered more was the posture: I refused to blame individuals, because the individuals were not the failure mode. The failure mode was a system that had been built without enough external adversarial input.
AI safety standards written by three labs face the same structural risk. They will be elegant. They will be internally coherent. And they will be optimized for a set of failure modes that the drafters were already thinking about — which is to say, the failure modes their own products exhibit.
Infrastructure and Compute: The Hidden Axis
There is a dimension to this story that the announcement completely omits, and it is the one that will ultimately determine whether the standards matter.
AI safety, at scale, is a compute problem. Red-teaming a frontier model costs millions. Running capability evaluations at the cadence required to catch emergent behavior requires dedicated inference clusters. Third-party auditing — the thing that would make a safety standard credible rather than self-certified — requires auditors who have access to the models, which requires contractual and infrastructure arrangements that most independent safety organizations cannot afford.
So here is the quiet implication: any safety standard that requires independent verification implicitly requires the incumbents to fund the verification infrastructure. And whoever funds the verification infrastructure has leverage over its conclusions.
This is precisely the dynamic I documented in the 2021 Axie Infinity contract forensics. When I coordinated with five other independent researchers to publish a joint threat assessment on the $SLP claim mechanism, our independence was possible precisely because we were unpaid and had no financial stake in the game. The moment that assessment would have required access to proprietary infrastructure, our independence would have been constrained by whoever controlled the access.
The same logic applies here. A safety standard with real teeth requires auditors with real access. Real access requires real infrastructure. Real infrastructure costs real money. And the entities with real money to spend on safety infrastructure are — structurally, necessarily — the same entities being audited.
The crypto industry solved a version of this problem with a concept it borrowed from political philosophy: separation of powers enforced by cryptographic guarantees. Consensus mechanisms work not because miners are virtuous, but because the protocol makes cheating more expensive than honesty. There is no equivalent mechanism in the current AI safety conversation. There is only trust — and trust, in the absence of verifiable mechanism, is just another word for concentration of control.
Contrarian: The Uncomfortable Blind Spots
Let me say the thing that will make some of my readers uncomfortable.
The problem with this collaboration is not that it is happening. The problem is that it is necessary, and the fact that it is necessary tells us something we should have already known: the AI safety field has not produced a credible independent governance layer, and the vacuum is being filled by the entities with the strongest incentive to define safety in a way that happens to fit them.
We should not be surprised by this. We should be angry at ourselves for not building the alternative sooner.
For the past three years, the crypto community — which has more experience with decentralized governance than any other industry on Earth, and most of that experience is scar tissue — has watched the AI discourse unfold as though it were a spectator sport. We have built DAOs, we have lived through governance attacks, we have watched token-voting plutocracies emerge, and we have developed genuine expertise in the failure modes of collective decision-making. And we have, for the most part, kept that expertise inside the crypto bubble while the most consequential governance question of our generation is decided by three corporate boards.
This is a failure of bridge-building, and I include myself in the indictment. The institutional-grassroots bridge I have tried to build in my own writing — translating cryptographic implementation for regulators, translating institutional priorities for builders — is exactly the bridge that is missing here. The AI safety conversation needs people who understand mechanism design, and the people who understand mechanism design are, disproportionately, in crypto. We were not in the room. We did not ask to be in the room. And the room was therefore defined without us.
There is a second blind spot, subtler and more dangerous. The framing of this collaboration as "safety" makes it very difficult to criticize, because criticizing it sounds like criticizing safety. This is a rhetorical structure, not a technical one, but it shapes what can be said in public. The same dynamic played out in DeFi during the Terra collapse: the algorithmic stablecoin was marketed as a mechanism for financial inclusion, and pointing out the rebalancing fragility sounded like opposing inclusion. The mechanism and the marketing were fused, and dissenting technical analysis was coded as opposing the value rather than questioning the implementation.
We should expect the same here. Any technical critique of these standards will be framed as opposing AI safety. The companies know this. It is, in the language of game theory, a dominant strategy: attach your commercial architecture to a value that no one can publicly reject.
And here is the third blind spot, the one that will matter most in twelve months. If these three labs succeed in defining the safety standard, they will also define the baseline against which every decentralized AI project is measured. Bittensor subnetworks, autonomous agent protocols, on-chain inference markets — none of these will be able to ignore the standard, because enterprise buyers and regulators will demand compliance. And none of them will be able to shape the standard, because they were not in the room. The decentralized AI thesis will not be killed by bankruptcy. It will be killed by compliance. Slowly, expensively, one audit at a time.
Takeaway
The interesting question is no longer whether Anthropic, OpenAI, and Google will produce a safety standard. They will. The interesting question is what the standard looks like when the first draft leaks, and who is allowed to comment on it before it becomes the baseline.
Watch for three signals over the next two quarters. First, the publication mechanism: does the draft get released for public comment, or does it appear as a finished framework with a press release? Second, the participation list: does Microsoft, Meta, or any non-Western lab get added, or does the original three remain the permanent drafting committee? And third, and most importantly, the verification architecture: does the standard require independent auditors with guaranteed access, funded by a neutral body — or does it rely on self-certification dressed in compliance language?
The answer to the third question will tell us whether this is safety or a moat. I have my priors. But I would love to be wrong.