Tracing the liquidity ghost in the machine — that is what I found myself doing when I read the second iteration of Anthropic's Responsible Scaling Policy (RSP) risk report, released in the quiet June of 2025. The crypto markets were euphoric, Bitcoin hovering near new highs, and the AI trade was the new narrative fuel for every token claiming to be the 'decentralized brain.' Yet here, in the dry prose of a safety framework, I saw a different liquidity story: the slow, invisible transfer of trust from open protocols to closed corporate governance. The report itself is not about blockchain, but its implications for the intersection of AI and crypto are profound. It signals that the control of frontier intelligence is consolidating into a handful of private actors, and the crypto industry, which dreams of decentralized AI agents, is sleepwalking into a dependency on these very same gatekeepers.
Context: The Architecture of Self-Regulation
Anthropic’s RSP, first published in May 2023, is a pioneering attempt to quantify and manage the catastrophic risks of large language models. It borrows the concept of biosafety levels (BSL) from microbiology, mapping model capabilities to four tiers: ASL-1 (minimal risk) through ASL-4 (near-AGI existential risk). The second report, released roughly two years after the first, confirms that the framework has moved from a static policy document to a dynamic, operational evaluation cycle. The core technical content involves assessing Claude 3.5 series models against benchmarks for CBRN (chemical, biological, radiological, nuclear) weapons proliferation, cyberattack capabilities, and autonomous replication. It is a methodological innovation in AI safety governance, not a breakthrough in algorithm design.
But for those of us who spent years in the cryptographic trenches—auditing zero-knowledge proofs, modeling liquidity flows in DeFi, and advising central banks on CBDC privacy architectures—the RSP’s structure rings a familiar bell. It is a permissioned system dressed in the language of safety. The ASL thresholds are set by Anthropic itself, the evaluations are conducted internally, and the resulting restrictions (weight access controls, KYC for model users, deployment limits) are enforced by the company without independent oversight. This is not a criticism of good intentions; it is an observation of power concentration. The report makes no mention of third-party audit, and the policy text only promises to 'plan for' external review. The self-evaluation, self-publication, self-supervision loop is the central structural weakness.
Core: The Crypto Convergence — Where AI Safety Meets On-Chain Trust
The crypto industry has been racing to integrate AI agents into DeFi, prediction markets, and autonomous organizations. The narrative is that blockchain provides trustless execution for AI decisions, with oracles verifying outputs and smart contracts enforcing rules. But the RSP report reveals a hidden dependency: the most capable AI models—those that could power meaningful autonomous agents—are likely to be classified as ASL-3 or higher, triggering strict access controls. Anthropic, like its competitors, will not open-source the weights of such models. The API access will be subject to KYC, usage monitoring, and potential revocation. This means that any crypto project seeking to integrate a frontier AI model will be building on a permissioned, revocable layer of centralized control.
From my experience at the intersection of cryptography and CBDC design, I have seen this pattern before. The Egyptian central bank’s digital pound prototype, which I advised on in 2023, included a zero-knowledge compliance layer that allowed transaction monitoring without revealing identities. The principle was that privacy could be eroded not by code, but by consensus—by the agreement of all parties to accept a surveillance layer in exchange for access. The same dynamic is unfolding in AI. The RSP’s graded access model means that the most powerful AI will be available only to approved entities, and those approvals will be controlled by a handful of corporations. The crypto ecosystem, which prides itself on permissionless innovation, will be forced to either accept second-tier models (open-source but less capable) or become a client of the very gatekeepers it sought to bypass.
The ETF wave washed away the retail tide — and now the institutional wave is washing away the trustless promise. I recall the liquidity analysis I performed during the Ethereum Merge in 2022, when I quantified how staking yields would synchronize with fiat liquidity cycles. The same macro lens applies here: the RSP is not just a safety document; it is a liquidity device. By establishing a credible self-regulatory framework, Anthropic lowers the political risk premium for institutional investors. The $184 billion valuation (as of mid-2024) and the multi-billion-dollar cloud deals with AWS and Google Cloud are built on the promise that the model will not cause a catastrophic event. The RSP is the insurance policy that allows investors to sleep at night. But the premium for that insurance is paid by the open ecosystem: the possibility that high-capability AI will remain locked inside corporate walls, accessible only via monitored APIs, while the decentralized alternative remains stuck at GPT-3.5 levels of competence.
Contrarian: The Decoupling Thesis — Is Safety Really a Crypto Problem?
The prevailing narrative in crypto circles is that AI safety is a separate concern, best left to the AI labs, while blockchain focuses on decentralization and value transfer. I argue the opposite: the RSP’s self-regulatory model is a direct threat to the crypto ethos, but it also presents a contrarian opportunity. The decoupling thesis—that crypto markets can ignore AI governance—is a dangerous blind spot. The liquidity of the entire AI+ crypto narrative depends on the ability to deploy autonomous agents on-chain with minimal permission. If the best models are only available via corporate APIs, then the 'decentralized AI' dream becomes a facade: the agent is just a smart contract that calls an API, and the API provider can revoke access at any time.
But here is the contrarian angle: the RSP’s structural weakness—its lack of external audit—is exactly where crypto can intervene. Zero-knowledge proofs can verify that a model is being used within certain safety constraints without revealing the model weights or the user’s identity. On-chain reputation systems can provide the KYC function without centralizing identity data. The very tools that crypto has developed for financial privacy and trustless verification can be applied to AI safety. History rhymes in the ledger — the same cryptographic primitives that enabled decentralized finance can enable decentralized AI safety attestation. The question is whether the AI labs will accept such external verification, or whether they will continue to rely on self-regulation.
Furthermore, the RSP’s focus on catastrophic risks (CBRN, cyberattacks) leaves a large blind spot in everyday social harms: bias, discrimination, privacy invasion. The crypto community, which has its own struggles with social consensus and governance, could fill this gap by developing frameworks for continuous ethical auditing of AI models on-chain. The Ethereum merge was a fever dream for liquidity, but the real merge—between AI and crypto—will require a fusion of safety paradigms, not just token narratives.
Takeaway: The Next Cycle — Will We Choose Permissioned Intelligence?
As I reflect on the RSP second report, I see the outlines of a new macro cycle. The first cycle of crypto was about peer-to-peer value transfer. The second cycle was about smart contracts and DeFi. The third cycle, now unfolding, is about the convergence of AI and blockchain. But the liquidity that powers this cycle flows through institutional channels, and those channels are gated by safety frameworks like the RSP. The choice is not between safety and freedom; it is between self-regulation and external audit; between permissioned intelligence and trustless verification. The crypto industry has a unique opportunity to architect the latter, but only if it recognizes that the ghost in the machine is not a technical bug—it is an institutional governance gap. We sleepwalk into a digital panopticon if we assume that safety can be left to the labs. The next bull run will be won by those who build the cryptographic infrastructure for AI safety attestation, not by those who simply add 'AI' to their token name. The liquidity is there, but the logic must be rebuilt.
