Hook: The Signal Without Substance
On June 2025, Anthropic released its second Responsible Scaling Policy (RSP) risk report. The document spans 47 pages, contains no model weights, no benchmark scores, and no independent audit signatures. Yet it is being hailed as a milestone in AI safety governance. As a DAO Governance Architect who has spent years designing verifiable, auditable decision-making systems for decentralized protocols, I see a familiar pattern: a beautifully structured framework that relies on the very entity it polices to enforce its own rules. In blockchain, we call this a "trust me" system. And we all know where that ends.
The report is a self-assessment of Claude 3/3.5 series models against ASL-2 to ASL-4 thresholds, focusing on CBRN, cyber capabilities, and autonomous replication. It claims to have transformed the RSP from a static document into a dynamic evaluation mechanism. But the question that matters for any governance system—whether it governs a DAO treasury or a frontier AI model—is not whether the rules are written, but whether they are enforced with integrity. And integrity requires independent verification. The community has been given a signal that the framework is running, but no information on whether the checks are real.
Context: The Governance Architecture of RSP
To understand why this matters, we must first parse the RSP's design. Anthropic borrowed the ASL (AI Safety Level) concept from biosafety's BSL framework. The levels run from 1 to 4, with ASL-3 being the critical threshold where model capabilities pose significant dual-use risks. The policy defines specific safeguards that must be triggered at each level: KYC for model access, weight access controls, vulnerability reporting mechanisms, and deployment restrictions. In theory, this is a sound governance structure—clear thresholds, defined actions, and escalation paths. It mirrors what we do in DAO governance when we set treasury withdrawal limits, quorum thresholds, and timelock delays.
But here is the foundational difference: in a well-designed DAO, the rules are encoded in smart contracts that execute autonomously. The code is the law. The community can audit the code, verify the parameters, and observe the outcomes on-chain. Anthropic's RSP, by contrast, is a document. The thresholds are determined by internal judgment calls. The enforcement depends on internal compliance teams. The audit trail is internal. There is no on-chain verification, no public validator set, no decentralized oracle that feeds the ASL level into a immutable decision engine. The system is opaque by design, and that opacity is its greatest weakness.
Core: Three Structural Faults in the RSP Governance Model
Fault 1: The Self-Assessment Circularity
The report evaluates whether Claude models meet ASL-3 criteria. The evaluation methodology is designed by Anthropic, executed by Anthropic employees, and published by Anthropic. This is a closed loop. In my experience auditing DAO governance proposals, any system that lacks external validation eventually develops a blind spot. The proponent becomes the judge. The incentive alignment is misaligned: if the model is found to be ASL-3, the company must impose costly restrictions. There is a natural pressure to set the bar high enough to avoid triggering those restrictions. The report does not disclose whether an independent third party was granted access to the evaluation data or the test sets. The policy text from earlier versions promised third-party audits, but there is no evidence in this report that such audits have occurred. Without that, the RSP is a self-report card with no teacher.
Fault 2: The Threshold Discretion Problem
ASL-3 thresholds are defined in terms of capability levels—for example, the ability to significantly lower the barrier to creating a bioweapon. How do you measure that objectively? The science of evaluating such capabilities is still nascent. Anthropic's internal team uses red-teaming and benchmark tests, but the choice of benchmarks, the scoring criteria, and the pass/fail lines are all proprietary. This gives Anthropic enormous discretion. In a DAO, if a governance proposal's success metric is ambiguous, we write it into the smart contract as a clearly defined condition. Here, the condition is a moving target. The report itself likely contains the results of these evaluations, but the article we are analyzing does not provide them. The hidden information is that the threshold setting is not just a technical decision; it is a strategic one. Setting the bar too low means admitting the model is dangerous. Setting it too high means the framework is toothless. Either way, the public cannot verify which is happening.
Fault 3: The Coverage Blind Spot
RSP focuses exclusively on catastrophic risks: CBRN, cyberattacks, autonomous replication. It does not cover the more common social harms: bias, discrimination, privacy violations, psychological manipulation. As a governance architect, I know that any framework that ignores the most probable risks while only addressing the extreme tail risks is a sign of selective governance. It is like a DeFi protocol that has a multi-sig for the treasury but no mechanism to prevent a flash loan attack on its liquidity pool. The protocol announces it is "secure" because the treasury is safe, while the users are being drained. RSP's narrow focus allows Anthropic to claim a high standard of safety without addressing the daily ethical risks that affect millions of users. This is not an oversight; it is a feature. Catastrophic risk management is a more defensible narrative for investors and regulators. Social risk management is messy, expensive, and often politically charged. By ignoring it, the RSP creates a governance gap that competitors like OpenAI or Google could exploit if they design a more comprehensive framework.
Contrarian: The Case for RSP as a Liability
Now, let me challenge the prevailing narrative. Many analysts view RSP as a competitive advantage for Anthropic, a trust-building asset. But from a governance perspective, a framework that is not independently verifiable is a liability waiting to materialize. The moment a public incident occurs—a Claude model being used to generate harmful content at scale, or a leak of model weights—the entire RSP edifice will be questioned. Was the model evaluated for that risk? If not, why not? If yes, why was it allowed to be deployed? The RSP creates a built-in expectation of safety. When that expectation is violated, the reputational damage is far greater than if no framework existed. It is the same reason why some DAOs avoid formal risk frameworks: once you publish a risk model, you are accountable for its failures. Anthropic has painted itself into a corner. It must now ensure that every future incident is within the scope of RSP, or the framework becomes a weapon against it.
Furthermore, the RSP's reliance on cloud infrastructure from AWS and Google Cloud introduces a second layer of trust dependency. The model weights are stored on cloud servers. The access controls are implemented by cloud providers. The security of the entire system depends on the integrity of both Anthropic and its cloud partners. In a decentralized context, we would never trust a single custodian with such critical assets. We would use a threshold signature scheme, distributed key management, and on-chain verification. Here, there is no such redundancy. If AWS suffers a breach, the RSP safeguards are compromised. The report does not address this dependency. It is a governance blind spot that traditional risk analysts would flag immediately.
Takeaway: The Real Test Will Be the ASL-4 Decision
Anthropic's RSP is a pioneering effort in AI self-governance. It is more structured than anything else in the industry. But as a governance architect, I judge systems by their ability to enforce hard constraints, not by their documentation. The real test will come when Anthropic develops a model that meets ASL-4 criteria—the threshold for extreme risk. At that point, the RSP mandates deployment restrictions that could halt the company's commercial trajectory. Will Anthropic honor the framework? Or will it find a way to reinterpret the thresholds? That decision will reveal whether the RSP is a genuine governance mechanism or a marketing document. The second report does not answer this question. It only tells us that the process is running. But running a process is not the same as building trust. Trust is earned through verification.
"Verify everything, trust nothing." That is the motto of every effective governance system. Anthropic has given us a framework to verify, but has not given us the keys to do so. Until independent auditors are given full access to the evaluation data, test sets, and incident logs, the RSP remains a promise. And in the world of decentralized governance, promises without execution are just noise. The AI industry needs to learn what blockchain already knows: code is the only law that holds. Until the RSP is encoded into a verifiable, auditable, and immutable system, it is just a PDF. And PDFs never stopped a rogue model from being deployed.
"Skepticism is the first line of defense." This report is a step forward, but the path to true AI governance is still long and requires a shift from self-regulation to distributed accountability. The community should demand transparency, not just reports. The next RSP update should include a public audit trail, a third-party validation report, and a clear explanation of the threshold calibration process. Otherwise, the governance architecture will remain a beautiful facade with no load-bearing walls.