Anthropic scored C+. OpenAI scored C. The industry average is barely passing. This is not a technical benchmark. It's a governance scorecard. The numbers land like a hammer on a keycap: blunt, unambiguous, and lacking nuance. As someone who has spent years dissecting smart contracts and zero-knowledge proofs, I recognize the pattern. The gap between public promises and on-chain reality is a familiar one. In crypto, we call it the audit gap. In AI, it's the safety index gap. The recent AI safety index, reported by an unnamed source, places two of the most influential AI companies in the same grade range as a mediocre student. But the media treats it as a verdict on model capability. It's not. It's a diagnosis of governance hygiene. And like any diagnosis, it demands scrutiny of the diagnostic tools themselves.
The article that triggered this analysis is a three-paragraph industry briefing. It states that Anthropic earns a C+, OpenAI a C, and that overall safety commitments are declining. It also flags the deepening ties between AI companies and the military as a concern. That's it. No methodology, no scoring dimensions, no historical comparison, no competitor list. The article is a signal, not a study. My job is to treat it as a forensic artifact—a piece of raw data that reveals more about the state of AI safety reporting than about the actual safety of these models.
Context: The Governance Scorecard
AI safety indexes are not technical benchmarks. They do not measure hallucination rates, jailbreak success probabilities, or data leakage vulnerabilities. Instead, they measure public commitments, transparency policies, red teaming practices, and external audit availability. Think of them as a compliance checklist for a heavily regulated industry that doesn't yet have clear regulations. This is exactly the world of cryptocurrency security audits. A smart contract audit, for example, produces a score or a list of findings. But the real safety of a DeFi protocol depends on the auditor's methodology, the code's attack surface, and the trust assumptions in the oracle or relay network. The AI safety index is no different. It's a proxy for governance, not a guarantee of safety.
Core: The Anatomy of a C+ Rating
Let me break down what a C+ rating means in engineering terms. It means that the company's safety practices are visible but incomplete. It's like a smart contract that passes a basic linting check but has no formal verification. It's like a wallet that supports multi-sig but distributes key shares without a threshold protocol. In my experience auditing custodial solutions for institutional ETFs, I saw exactly this pattern: companies would claim multi-party computation, but the key-shares distribution logic had a single point of failure. The C+ rating says the same for AI: there is a safety process, but it's not robust.
Anthropic vs OpenAI: The Strategic Divergence
Anthropic has long positioned itself as the "safety-first" AI company. Its CEO has testified before Congress, its published papers emphasize constitutional AI, and its public relations consistently leads with caution. This branding certainly influences a governance score. OpenAI, by contrast, has been product-first: ChatGPT, GPT-4, and the API ecosystem drive its narrative. The C+ vs C gap is small, but it reflects a real difference in how each company presents its safety posture. However, a C+ is not a passing grade. In any academic system, C+ is below average. The fact that the industry's safety leader earns a C+ tells us that the bar is low. The industry's safety leader is not safe—it's just less unsafe than the others.
Military Ties: The Unspoken Vulnerability
Both companies have deepened relationships with the U.S. Department of Defense. The article flags this as a concern. From a cryptographic perspective, this is a trust assumption. When a model is used in a military context, the threat model changes. The attacker is not a script kiddie—it's a state actor. The safety standards required for a battlefield assistant are different from those for a chatbot. In crypto, we say that code is law, but bugs are reality. In AI, the code is the model weights, and the reality is that military contracts introduce a new set of attack surfaces. The current safety index does not account for this. It treats all deployment contexts as equal. That's a flaw.
The Missing Verifiability
Here is the core insight that the article misses: the AI safety index is not cryptographically verifiable. The scores are based on self-reported data and expert assessments. There is no proof that the red teaming exercises actually happened, or that the alignment training was applied correctly. In crypto, we have solved this problem for certain domains. Zero-knowledge proofs allow one party to convince another that a computation was performed correctly without revealing the inputs. I built a prototype of this for AI model outputs in 2026—a ZK-circuit that proves an inference was generated without tampering, using a specific dataset and model weights. The same principle can be applied to safety audits. Imagine an AI company publishing a zk-proof that its red teaming results were generated by a specific set of testers, or that its alignment fine-tuning used a particular dataset. That would be a verifiable governance signal. The current safety index is just a claim. And in engineering, claims without proof are vulnerabilities.
Institutional Adoption and the Audit Gap
In early 2024, I audited the custodial wallet solutions used by major asset managers for their spot Bitcoin ETFs. The findings were sobering. The multi-sig threshold logic had critical gaps. The key-shares distribution protocol was not actually secure—it relied on a single trusted dealer. The companies marketed "institutional-grade security" but the implementation was amateur. The same dynamic is playing out in AI. The safety index is a marketing tool until it's not. Financial institutions, healthcare providers, and government agencies will eventually demand more than a letter grade. They will demand evidence. They will demand cryptographic audits. The companies that invest in verifiable safety protocols will have a competitive advantage. The ones that rely on C+ ratings will be left behind.
Contrarian: The Low Score Might Be a Honest Signal
There is a counter-intuitive take here. Perhaps the low scores are a sign of transparency. A company that admits it earns a C+ is more credible than a company that claims an A+ but hides its methodology. In crypto, we have learned that the most dangerous protocols are the ones that pass audits but fail in practice. The Terra/LUNA crash was a case in point. The Anchor Protocol had been audited. The code was clean. But the economic model had a fatal flaw that no audit caught. The AI safety index might be similar: it measures governance, but not the actual safety of the model. A C+ rating might indicate that the company is honest about its limitations. That is a feature, not a bug. Privacy is a feature, not a bug—and so is transparency. The real danger is when companies claim perfect safety but have no verifiable proof. That is the path to the next LUNA-like collapse.
Takeaway: The Need for Verifiable Governance
The AI safety index is a symptom of a larger problem: the industry lacks a universal standard for verifiable trust. Until AI companies adopt cryptographic proofs for their safety claims, ratings will remain opinions. Math doesn't negotiate. The industry needs to move from governance scores to verifiable audits. The blockchain community is already building the infrastructure for this. The question is whether AI companies will adopt it. The C+ and C grades are not a verdict. They are a call to action. The next step is not to improve the score—it's to change the scoring system.
Code is law, but bugs are reality. The AI safety index is a bug report. It's up to the industry to fix the code.