Anthropic’s new AI model, Claude Mythos Preview, has autonomously discovered critical weaknesses in TLS, AES-GCM, and SSH cryptographic protocols — the same infrastructure securing wallets, bridges, and smart contracts holding over $200 billion in decentralized finance assets. The announcement arrived April 7 alongside Project Glasswing, a restricted-access security initiative with 12 institutional partners including AWS, Apple, Google, Microsoft, and JPMorganChase.
Decentralized finance, or DeFi, refers to financial services built on public blockchains — lending, trading, and yield-earning platforms that operate without banks or intermediaries, relying entirely on code and cryptographic security.
- Anthropic’s Mythos Preview produced 181 working exploits against Firefox 147’s JavaScript engine autonomously — compared to 2 for predecessor Opus 4.6.
- The model found weaknesses in TLS, AES-GCM, and SSH, the cryptographic protocols securing DeFi wallet connections, oracle feeds, and validator communications.
- Anthropic warned that friction-based DeFi defenses — multisig approvals, timelocks, and audit certifications — may no longer hold against AI-speed attackers.
- Access to Mythos Preview is currently restricted to 40+ critical infrastructure organizations through Project Glasswing. A broader release timeline has not been announced.
Published: April 9, 2026 09:00 UTC
What Mythos does differently
Existing vulnerability scanners run pattern-matching against known flaws. Mythos works like a skilled human researcher, but faster and cheaper. The model reads source code, forms hypotheses, executes software, employs debuggers, and generates working proof-of-concept exploits without prompting.
The scale of what it found is hard to ignore. Mythos identified a 27-year-old denial-of-service flaw in OpenBSD, a 16-year-old integer overflow in FFmpeg’s H.264 codec that five million automated test iterations had missed, and a 17-year-old FreeBSD remote code execution vulnerability granting unauthenticated root access. Third-party reviewers evaluated 198 of Mythos’s findings: 89% received the same severity rating as human assessors; 98% were within one rating level.
Cost is the part that changes the threat model. Mythos found the OpenBSD flaw for under $20,000. An exploit chain targeting patched Linux kernel CVEs was completed in under one day for less than $2,000. These figures represent a drop of several orders of magnitude compared to traditional security research.
Anthropic is providing $100 million in model usage credits and $4 million in donations to open-source security organizations as part of the initiative, according to Project Glasswing’s official page.
The DeFi exposure
DeFi protocols use two security layers. The first is cryptographic infrastructure — protocols like TLS, AES-GCM, and SSH that encrypt data moving between users, nodes, validators, and price oracles. The second is friction-based governance: multisig approvals, timelocks, and audit reports that make attacks slow and detectable.
Multisig governance requires multiple parties to sign off before a blockchain transaction executes. A timelock delays execution of on-chain actions by a set period, giving communities time to catch malicious proposals before they go live.
Anthropic stated in its announcement: “mitigations whose security value comes primarily from friction rather than hard barriers may become considerably weaker against model-assisted adversaries.” Multisig and timelocks are explicitly friction-based. They slow attacks. Against a model operating at machine speed against cryptographic vulnerabilities, slowing is not the same as stopping.
Crypto protocols lost over $1.8 billion to hacks and exploits in 2025, according to DeFiLlama data. Those attacks were human-speed. The risk calculus shifts considerably if attackers gain access to tools with Mythos-level capabilities, as TechCrunch reported when covering the model’s debut.
What cryptographic protocols are actually at risk
TLS is a security protocol that encrypts data transmitted between applications and servers, including API communication between crypto wallets and blockchain nodes. AES-GCM is a symmetric encryption algorithm used to secure data at rest and in transit. SSH is the remote access protocol used to manage server infrastructure.
DeFi infrastructure depends on all three for wallet API calls, validator node connections, and off-chain price oracle feeds. A verified TLS vulnerability could enable a man-in-the-middle attack on oracle data — the price feeds that DeFi lending protocols use to decide when to liquidate collateral. A compromised oracle is one of the most common attack vectors in DeFi, and has been responsible for hundreds of millions in losses over the past three years.
The implications are not hypothetical. As Help Net Security reported, Mythos successfully built privilege escalation exploits for more than 50% of 40 filtered Linux kernel CVEs from 2024 to 2025. These are patched vulnerabilities — and Mythos cracked them anyway.
For broader context on how AI is reshaping crypto infrastructure, see the recent Circle CEO piece on crypto as an economic OS for the AI era.
What comes next
Mythos Preview is not publicly available. Project Glasswing partners — which include 40+ critical infrastructure organizations beyond the 12 named launch partners — are using the model defensively: finding and patching flaws before broader release.
Anthropic has not announced a timeline for public deployment. But the window between restricted access and general availability tends to compress quickly in AI development. The Axios analysis of Mythos framed the current period as the only available window for defenders.
DeFi protocols with significant treasuries or reliance on standard internet cryptographic infrastructure should treat this as a prompt to audit off-chain systems now. Smart contract code alone is not the full attack surface. Validator security, oracle integrations, and API gateway configurations all sit behind the same TLS and SSH infrastructure Mythos has already demonstrated it can probe.
Protocols that have faced security incidents in the past are worth watching for incident disclosures in the coming weeks. For reference on how regulators have approached prior crypto security failures, see our coverage of the FBI’s seizure of the Tickmilleas scam site.
Frequently asked questions
What is Project Glasswing?
Project Glasswing is Anthropic’s initiative to share its Mythos AI model with vetted critical infrastructure organizations before any public release. Twelve named launch partners include AWS, Apple, Google, Microsoft, and JPMorganChase, with over 40 total organizations receiving access. Anthropic committed $100 million in model usage credits and $4 million in open-source security donations to the program.
How does Mythos AI threaten DeFi specifically?
Mythos found weaknesses in TLS, AES-GCM, and SSH — protocols that protect wallet connections, oracle data feeds, and validator communications in DeFi. It also operates at machine speed and dramatically lower cost than human researchers, which weakens friction-based defenses like multisig approvals and timelocks that are designed to slow attacks rather than block them at the code level.
Is DeFi at immediate risk from Mythos?
Not immediately. Mythos Preview is restricted to Project Glasswing partners and is being used defensively. However, Anthropic explicitly warned that friction-based security measures may weaken significantly as AI capabilities become more accessible. DeFi teams are advised to audit off-chain infrastructure — oracle integrations, validator nodes, and API configurations — before broader AI deployment changes the threat environment.
