The model name is wrong. That's where any credible analysis of Payward's Project Glasswing announcement should begin—and end. "Claude Mythos 5" does not exist. Anthropic's current lineup stops at Claude 3.5, 3.7 Sonnet, and Claude 4. There is no Mythos variant. Not in their documentation, not in their API, not in any public research paper. The math didn't check out before I even opened the press release.
This is not a trivial typo. In an industry where every codebase is a fortress and every vulnerability is a potential $100 million exploit, naming matters. If the core technology is misrepresented, the entire narrative is suspect. I've spent 13 years dissecting projects that claim to revolutionize security, from the ICO era's whitepaper fallacies to the DeFi Summer's un-audited smart contracts. The first lesson: never trust a headline that can't get the product name right.
Context: The AI Security Hype Cycle
Payward, the parent company of Kraken, announced its participation in Anthropic's Project Glasswing—an initiative described as a pilot program for AI-driven vulnerability discovery. The pitch is standard: use large language models to proactively scan software for bugs, reducing the risk of exploits that have cost the crypto industry over $2.5 billion in bridge hacks alone. Kraken, a 14-year-old exchange with a relatively clean security record, is positioning itself as a forward-thinking guardian of digital assets.
This fits the 2024-2025 market narrative. AI + Crypto security is a hot vertical. Startups like Socket, Lasso Security, and even Google's LLM-powered vulnerability research have been making headlines. The idea is sound: LLMs can analyze code patterns faster than humans, flagging potential issues before they become headlines. But sound ideas and sound implementations are separated by a chasm of engineering rigor.
Core: A Systematic Teardown of the Announcement
I've audited dozens of security tools—from static analysis frameworks to formal verification suites. The minimum viable disclosure for any serious vulnerability discovery system includes: methodology, false positive rate, detection rate, integration complexity, and validation process. The Payward-Anthropic announcement provides none of these. Here's what we know:
- The system uses "Claude Mythos 5" to search for software vulnerabilities.
- It is part of Project Glasswing, an Anthropic pilot.
- Kraken expects AI to play a proactive role in cybersecurity.
That's the entire data set. No technical paper. No benchmark. No case study. No mention of how the model is fine-tuned, what prompt engineering techniques are used, or how it integrates with existing CI/CD pipelines. As an analyst, this is not a security tool—it's a press release.
The Model Problem
Let me be direct: if the model name is fabricated, the entire project may be fabricated. But even if we assume a translation error or internal codename, the lack of verifiable details is a red flag. I've seen this pattern before. In 2020, I traced the Harvest Finance exploit to an unencrypted documentation that promised a "proprietary risk engine"—turns out it was a simple if-else statement. When technical details are withheld, it's usually because they don't exist.
Furthermore, LLMs in code auditing are not a solved problem. They suffer from high false positive rates—sometimes 30-50%—which can lead to alert fatigue. More critically, they hallucinate vulnerabilities that don't exist, wasting developer time. Security isn't a feature you can bolt on with a chatbot; it's the foundation of the entire system. The foundation here is unverified.
Data Security Risks
Kraken's core code is its crown jewel. Sending that code to a third-party API, even with encryption, introduces supply chain risk. If Anthropic's model is trained on Kraken's codebase, that data becomes part of the model's weights—potentially extractable by future adversaries. I've consulted on institutional custody solutions where NDAs and air-gapped environments were mandatory. This announcement mentions none of those safeguards.
False Positives vs. False Negatives
The real danger isn't false positives—it's false negatives. If the AI misses a critical vulnerability, the team may assume the code is clean. Every rug has a seam you missed, and an AI that doesn't find it only provides a false sense of security. In my analysis of the Terra/Luna collapse, the same pattern emerged: a system that was believed to be stable because the models said so. Models are not oracles.
Contrarian: What the Bulls Got Right
Now, let me be fair. The bulls would argue that adopting AI security tools is a net positive. Kraken is a compliance-first exchange, and this move aligns with regulatory expectations from the SEC and CFTC for robust cybersecurity frameworks. The partnership with Anthropic—a company known for "responsible AI"—suggests Kraken is thinking about governance and ethics. The market is rewarding AI narratives, and this could indirectly boost trust in Kraken's brand.
I can't argue with the trend. AI will eventually transform security. But the gap between "eventually" and "now" is where most projects fail. The contrarian insight is that Kraken's decision is rational from a PR perspective, not a technical one. They are buying a seat at the table, not a solution. If they can later disclose concrete results—say, "Project Glasswing identified 3 critical vulnerabilities in our API"—this announcement becomes a foundation. Until then, it's speculation disguised as progress.
Takeaway: Demand Accountability
Payward and Anthropic must release the following: a technical whitepaper describing the model's architecture, a benchmark comparing it to existing tools like Semgrep or Snyk, and a data protection policy. Without these, the announcement is a hollow narrative. Hype burns out; structural integrity remains. The crypto industry has been burned by too many promises of silver bullets. This one is no different.
If you're a Kraken user, you should ask: what is the model's false positive rate? If you're an investor, remember that this has zero impact on any token price. Risk is not eliminated by ignoring it. The only way to verify this claim is to wait for the audit results—the real ones, not the AI-generated ones.
I will be watching. And I will be checking the math.