Anthropic’s Mythos AI model has found critical flaws in the cryptography libraries that underpin most of the internet’s secure infrastructure. Security researchers say DeFi protocols holding roughly $94 billion in user funds are acutely exposed. The warning arrived with unusual urgency: Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell convened an emergency meeting on Tuesday at Treasury headquarters with the CEOs of Citigroup, Morgan Stanley, Bank of America, Wells Fargo, and Goldman Sachs to discuss the threat posed by Mythos and models like it.
Decentralized finance, or DeFi, refers to financial applications built on public blockchains that execute transactions automatically through self-executing smart contract code, without banks or brokers as intermediaries.
- Mythos Preview found vulnerabilities in TLS, AES-GCM, and SSH cryptography libraries, including the same libraries securing bridges, wallets, and DeFi front-ends, with a certificate authentication bypass in the Botan library.
- The model built working privilege escalation exploits for over half of a set of filtered Linux kernel CVEs in under one day at a cost under $2,000, compared to the months and millions previously required.
- DeFi’s primary defenses (multisig governance and timelocks) are friction-based, not code-level barriers, and are now considered insufficient against AI-assisted attackers.
- Approximately $94 billion in total value is locked across DeFi protocols as of April 2026, with roughly $53 billion on Ethereum alone.
Published: April 10, 2026, Midday UTC
What Mythos can actually do
Anthropic’s Mythos Preview, unveiled April 7 as part of a restricted security initiative called Project Glasswing, is not a typical AI model. It operates using an “agentic scaffold”: isolated containers run target codebases while the model autonomously reads source code, forms hypotheses, runs debuggers, and produces proof-of-concept exploit reports without human guidance at each step.
In benchmark testing, Mythos produced 181 working Firefox exploits versus 2 for Anthropic’s previous Opus 4.6 model. It found a 27-year-old denial-of-service flaw in OpenBSD’s TCP stack and a 17-year-old remote code execution bug in FreeBSD’s NFS server, with findings matching professional contractor severity assessments 89% of the time. The economic shift is the operative fact: capabilities that previously required months and millions now cost under $2,000 and less than a day.
For DeFi, the most pointed findings involve cryptography. Mythos identified vulnerabilities in implementations of TLS, AES-GCM, and SSH: core protocols securing encrypted communications and authentication. A specific flaw was a certificate authentication bypass in the Botan cryptography library. These libraries are used by wallet providers, cross-chain bridges, and the web interfaces through which users interact with DeFi protocols.
Why DeFi is particularly exposed
DeFi protocols are open-source software, which creates both a strength and a liability. Human auditors and automated scanners have reviewed the code, but Anthropic’s own assessment notes that “mitigations whose security value comes primarily from friction rather than hard barriers may become considerably weaker against model-assisted adversaries.”
That description fits DeFi’s most common security controls precisely. Multisig governance requires multiple parties to approve on-chain transactions. Timelocks delay transactions for a set period to allow intervention. Both slow things down. Neither blocks an attack at the code level. A sufficiently capable AI that can identify and chain zero-day vulnerabilities in the libraries sitting beneath those controls renders them insufficient.
Cybersecurity researcher Marin Ivezic, who spent three decades in offensive security, put the asymmetry plainly: “The defender must find and fix every vulnerability. The attacker only needs to find one.” With Mythos compressing that search from months to hours, the balance shifts.
Cross-chain bridges, which have already lost over $2.5 billion to exploits since 2021 according to DeFi Llama, rely heavily on cryptographic validation of off-chain data. A model that can autonomously identify authentication bypasses in TLS or SSH implementations is particularly well-positioned to target bridge infrastructure.
The regulatory response
The Bessent-Powell meeting on April 8 was notable for its composition and its speed. Five of the six largest U.S. banks by assets sent their chief executives; JPMorgan CEO Jamie Dimon was absent. The stated goal was to ensure financial institutions are “aware of possible future risks” from Mythos and models with comparable capabilities, and are taking precautions. The meeting came one day before Anthropic publicly announced that Amazon, Apple, and JPMorgan (Dimon’s firm) were among the roughly 40 technology companies receiving restricted access under Project Glasswing.
The SEC is separately scheduled to hold a roundtable on April 16 to discuss the CLARITY Act, proposed legislation that would resolve regulatory jurisdiction over digital assets between the SEC and CFTC. No public link has been drawn between the Mythos threat and that session, but the timing means crypto firms may face simultaneous pressure from two directions: potential structural security vulnerabilities and incoming regulatory reclassification.
Project Glasswing’s model is essentially a defensive head start. Partner firms get restricted access to Mythos for “defensive security work” before the model or comparable tools become more widely accessible. The implication for DeFi protocol teams that are not among those 40 firms is that the gap between attacker capability and defender awareness is already open.
What protocol teams can do now
Security researchers highlight three near-term priorities for DeFi teams: audit cryptography library dependencies, specifically TLS, SSH, and AES-GCM implementations and patch known CVEs. Protocol teams should move security controls from friction-based mechanisms toward hard code-level barriers where possible and treat existing audit reports as a baseline rather than a ceiling. Mythos-class models have already identified vulnerabilities that human auditors and automated scanners missed. Protocol teams should expect adversarial AI to reach open-source targets before defenders are aware.
Frequently asked questions
What is Project Glasswing?
Project Glasswing is Anthropic’s controlled security initiative through which it is distributing restricted access to its Mythos AI model to approximately 40 partner organizations, including Amazon, Apple, and JPMorgan. The goal is to allow those firms to use the model for defensive security work and to strengthen their systems before similar AI capabilities become broadly available.
Are DeFi protocols already being attacked using Mythos?
No confirmed attacks using Mythos have been publicly reported as of April 10, 2026. The current threat is prospective: Mythos has demonstrated capabilities that, if applied offensively, would outpace the friction-based defenses most DeFi protocols rely on. The concern is that similar capabilities will emerge from other developers before all vulnerable systems are patched.
What is a smart contract, and why does cryptography matter for it?
A smart contract is a program stored on a blockchain that automatically executes transactions when predefined conditions are met , no bank or intermediary required. Cryptography secures the communications, authentication, and front-end interfaces surrounding those contracts. Vulnerabilities in cryptography libraries do not necessarily mean the contract code itself is compromised, but they can expose the access points users rely on to interact with it.
