Skip to main content

Project Glasswing: How Anthropic's $100M AI Security Cartel Forces Crypto Into a Two-Tier Defense Economy

· 10 min read
Dora Noda
Dora Noda
Software Engineer

On April 7, 2026, Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell pulled the CEOs of Citigroup, Morgan Stanley, Bank of America, Wells Fargo, and Goldman Sachs into an emergency meeting at Treasury headquarters. The subject was not a bank failure, a rate decision, or a sanctions regime. It was a single AI model built by a San Francisco research lab — Anthropic's Claude Mythos Preview — that had quietly found thousands of high-severity vulnerabilities in every major operating system and every major web browser, more than 99% of them still unpatched.

Three days earlier, Anthropic had announced Project Glasswing: a commitment of up to $100M in Mythos usage credits to a closed coalition of twelve technology, security, and financial giants — AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks — plus over 40 critical open-source maintainers. Everyone else, including Coinbase and Binance, was left to negotiate from outside the perimeter.

For crypto, the implications cut deeper than a typical security-tool launch. Glasswing is the first time a private AI lab has effectively defined a two-tier vulnerability-discovery economy, and the crypto industry — which lost over $3B to exploits in H1 2025 alone — has to decide whether it belongs on the inside or the outside of that perimeter.

What Mythos Actually Does​

Anthropic's own framing is unusually stark. In internal tests, Mythos identified a 27-year-old bug in OpenBSD that no human auditor had ever surfaced, then chained consecutive vulnerabilities to break out of modern browser sandboxes. Traditional smart contract audits take weeks. Mythos generates effective attack paths in seconds.

That asymmetry is the story. The model does not just flag candidate bugs; it auto-generates working exploit code and orchestrates multi-stage attack chains. Anthropic deemed the capability "super dangerous" for unsupervised public release, which is why Mythos Preview is not available via normal API access. Instead, it lives behind the Glasswing gate.

The coalition is not a research collaboration in the academic sense. Participants receive live access to Mythos to hunt vulnerabilities in their own systems — TLS implementations, AES-GCM primitives, SSH daemons, kernel code, and in JPMorgan's case, the internal payment and trading stacks that clear trillions of dollars daily. Anthropic has committed to publish a 90-day public report in early July 2026 summarizing what Glasswing has fixed.

Why Coinbase and Binance Are Now Negotiating From Outside the Wall​

Coinbase's chief security officer Philip Martin has publicly confirmed the company is in "close communication" with Anthropic, framing the objective as building an "AI immune system" — using Mythos defensively to scan its own systems before someone with a comparable capability uses it offensively. Binance's CSO described a parallel evaluation, citing both the defensive upside and the threat surface.

The asymmetry problem for crypto exchanges is brutal. A centralized exchange holds hot wallet keys, user balances, and a custody stack that any moderately motivated offensive operator would pay seven figures to probe. If Mythos — or a model of equivalent capability leaked from an employee, a state-sponsored actor, or an eventual open-weight competitor — ends up in attacker hands before exchanges harden their systems, the exploit window is measured in hours, not quarters.

That is the core of the Glasswing dilemma. Exchanges that are not inside the coalition cannot use Mythos to pre-audit their own code. They can use second-tier tools, but the capability gap matters. A bug that Mythos catches in 30 seconds might take a human auditor three weeks, and might be found by an adversary with comparable AI access in minutes.

The $3B Context: Why Speed Asymmetry Is an Existential Threat for DeFi​

H1 2025 saw over $3B in Web3 platform losses. Access control exploits alone accounted for $1.63B — the leading category in that period's OWASP Smart Contract Top 10. FailSafe's 2025 report tallied $2.6B in losses across 192 incidents. Immunefi has paid out over $115M in bug bounties across 400+ protocols and claims to have prevented more than $25B in potential losses.

Now overlay Mythos-class capability on that threat model. A protocol with $500M TVL that relies on a quarterly audit from a top-tier firm was already losing the race against well-resourced attackers. When one side of the table can auto-generate exploit chains in seconds, the audit cadence that defined DeFi security from 2020 through 2025 stops working.

The defensive equivalent exists but lags. CertiK's AI Auditor, open-sourced after six months of internal testing, achieves an 88.6% cumulative hit rate across 35 real 2026 web3 security incidents. It runs parallel specialized scanners through a multi-stage validator to filter duplicates and non-exploitable findings. CertiK has flagged over 180,000 vulnerabilities across its eight-year history and secured more than $600B in digital assets.

But 88.6% is not 100%, and an open-source auditor that runs in minutes is not the same as a frontier model that reasons about novel vulnerability classes in seconds. The gap between what Glasswing partners get and what public tools deliver is structural.

Three Competing Security Architectures​

The crypto industry now has to choose among three incompatible models for AI-era security:

Public bug bounties (Immunefi). Decentralized, economically aligned, proven at scale — $115M paid out, $25B saved. But the incentive structure assumes attackers and defenders operate at roughly equivalent speed. Mythos breaks that assumption. A white-hat researcher chasing a $50K bounty cannot outbid a state-sponsored actor paying $5M for a zero-day on a $10B protocol.

Open-source AI auditing (CertiK, Sherlock, Cyfrin). Democratic access to mid-tier AI capability, 88.6% hit rate, integrates into developer workflows. Preserves the crypto-native ethos that security tooling should be public. But the capability ceiling is below what Glasswing partners get, and the gap compounds as frontier models improve.

Gated-access frontier AI (Glasswing). Best-in-class vulnerability discovery, but only for members of a private coalition that currently does not include any crypto-native company. Creates clear tiers of cyber defense where the inside of the wall is safer than the outside.

The three models are not mutually exclusive — an exchange could run CertiK's auditor on every contract deployment, maintain an Immunefi bounty, and lobby for Glasswing partnership — but they imply very different industry structures. If Glasswing becomes the default tier for "systemically important" infrastructure, crypto's largest custodians face pressure to get in, and the protocols that can't get in face a pricing penalty on their risk premium.

The Systemic Framing Changes Everything​

What made the April 7 Bessent-Powell meeting remarkable is not the fact that regulators talked to bank CEOs about cyber risk. That happens routinely. The remarkable fact is the framing: AI-class cyber capability is now being treated as a potential catalyst for systemic financial events, on par with a sovereign debt crisis or a major clearinghouse failure.

That framing has second-order consequences for crypto. Stablecoin issuers holding tens of billions in reserves, custodians holding institutional BTC and ETH, and the exchange matching engines that process hundreds of billions in monthly volume all sit squarely inside the definition of "systemically important" that regulators are starting to apply to AI cyber risk. If the next Powell-Bessent-style meeting happens and crypto leadership is not at the table, that is both a signal and a problem.

The regulatory signal matters because Glasswing's 90-day public report in July 2026 will publish both what partners fixed and what the broader industry should learn. If that report documents classes of vulnerabilities that Mythos found in critical infrastructure, and crypto protocols have not done equivalent work, the gap will be visible to regulators, insurers, and institutional allocators pricing counterparty risk.

What This Means for Infrastructure Providers​

Machine-speed offensive AI changes the audit cadence required to defend production systems. A protocol or infrastructure provider that relied on annual audits, quarterly pen tests, and reactive incident response needs to shift to continuous AI-assisted red-teaming. That is expensive, and the expense lands unevenly across the stack.

For RPC providers, API infrastructure, and node services that sit between agents and chains, the pressure is to harden the surface where machine-initiated traffic terminates. Agent-driven transaction volume already creates a different threat profile than human-driven dApps: burst-heavy, predictable schedules, and deterministic call graphs that an attacker can model more precisely than a dispersed human user base.

BlockEden.xyz operates enterprise-grade RPC and API infrastructure across Sui, Aptos, Ethereum, Solana, and other major chains, with security and reliability built to serve both human developers and autonomous agent workloads. Explore our services to build on infrastructure designed to hold up in an AI-accelerated threat environment.

The Open Question Heading Into July 2026​

The 90-day Glasswing report is the pivot. If it documents a large backlog of serious vulnerabilities fixed in AWS, Google, Microsoft, Apple, and JPMorgan systems, the case for expanding the coalition gets stronger, and pressure builds on Anthropic to add crypto-native members or to license Mythos-equivalent access through a formal vendor relationship. If the report underdelivers — overcounts CVE findings, documents mostly low-severity bugs, or surfaces issues that existing scanners already caught — the Glasswing model loses some of its regulatory mystique and the crypto industry's open-source alternative looks relatively stronger.

Either way, the status quo from 2020-2025 is gone. The combination of an emergency Bessent-Powell meeting, a $100M Anthropic commitment, a 99%+ unpatched rate on Mythos-discovered bugs, and $3B in annual DeFi losses means that AI-era security is no longer a research question. It is a market structure question, and crypto's answer will define whether the next $100B of on-chain value sits inside a defensible perimeter or outside one.

Sources​

Share on Twitter