Anthropic revealed Claude Mythos Preview on Apr. 7, 2026, as the most powerful AI model it has ever built, and the first it has explicitly refused to release to the public due to its capacity to discover thousands of zero-day software vulnerabilities across every major operating system.
The cryptocurrency industry, which lost a record 3.3 billion dollars to hacks in 2025, now faces the prospect that AI-powered offense could accelerate the very attack chains that already drain billions from exchanges, bridges and wallets each year.
TL;DR
- Anthropic's Claude Mythos Preview discovered thousands of zero-day vulnerabilities and was deemed too dangerous for public release, prompting the launch of a 104 million dollar defensive initiative called Project Glasswing.
- The crypto industry lost 3.3 billion dollars to hacks in 2025, with access control failures and supply chain attacks causing the majority of losses, precisely the types of multi-step attacks Mythos can plan and execute.
- Larger exchanges are investing aggressively in AI-powered security, while smaller DeFi protocols risk falling further behind, potentially accelerating consolidation across the industry.
What Claude Mythos Actually Is
Claude Mythos Preview first surfaced on Mar. 26, 2026, when Fortune discovered a draft blog post in an unsecured data cache connected to Anthropic's content management system.
Nearly 3,000 unpublished assets were exposed through a CMS configuration error. Anthropic confirmed the model's existence and called it a step change in capabilities.
The formal announcement arrived on Apr. 7 alongside a 244-page System Card. That document is the most detailed safety disclosure Anthropic has ever published.
The company stated plainly that it does not plan to make Claude Mythos Preview generally available.
The reason is cybersecurity. Mythos discovered thousands of zero-day vulnerabilities in every major operating system and web browser, including bugs that survived decades of human review.
A 27-year-old flaw in OpenBSD, an operating system specifically hardened for security, allowed remote crash of any machine. A 16-year-old vulnerability in FFmpeg was found in a line of code that automated testing tools had hit 5 million times without detection.
What alarmed researchers most is that these cyber capabilities were not specifically trained. They emerged as a downstream consequence of general improvements in code, reasoning and autonomy. That finding carries enormous weight for the AI industry.
As all frontier models improve at coding and reasoning, comparable offensive capabilities may appear within months.
Alex Stamos, former Facebook security chief, warned that open-weight models could reach similar capability levels in roughly six months.
Anthropic activated ASL-3 protections for Claude Opus 4 in May 2025, initially for chemical and biological risk. Multiple analysts indicate Mythos's cyber capabilities placed it at or near the ASL-3 threshold for cybersecurity as well, requiring enhanced safeguards.
Rather than release the model, Anthropic launched Project Glasswing, a 104 million dollar defensive cybersecurity initiative.
The name references a butterfly species with transparent wings. The initiative includes 100 million dollars in Mythos usage credits for partner organizations, 2.5 million dollars to Alpha-Omega and OpenSSF through the Linux Foundation, and 1.5 million dollars to the Apache Software Foundation.
The 12 founding partners span the technology industry:
- Amazon Web Services, Apple, Google and Microsoft on the cloud and operating system side
- CrowdStrike, Cisco, Broadcom and Palo Alto Networks for enterprise security
- NVIDIA for compute infrastructure
- JPMorganChase as the sole financial institution
- Linux Foundation and Anthropic itself
More than 40 additional organizations have joined since the launch. Microsoft reported substantial improvements over previous models on its CTI-REALM benchmark. CrowdStrike confirmed that frontier AI capabilities compound when paired with real-world threat intelligence. In 198 manually reviewed vulnerability reports, Mythos achieved 89 percent exact agreement with professional security contractors on severity assessment.
Also Read: Polkadot Bridge Exploit Lets Attacker Mint 1B DOT Tokens On Ethereum
Why Crypto Is Unusually Exposed to Mythos-Style Risk
Cryptocurrency is not just another software industry. It is one of the few sectors where software risk, financial risk and market risk are fused into the same system. A vulnerability in a traditional web application might leak data. A vulnerability in a DeFi protocol can drain hundreds of millions of dollars in minutes.
Several structural features make crypto uniquely exposed:
- Markets operate 24/7 with no circuit breakers, meaning exploits can be executed at 3 a.m. on a Sunday when response teams are offline
- Settlement is irreversible on-chain, so stolen funds cannot be clawed back through a bank's fraud department
- Wallet permissions grant direct control over assets, and a single compromised private key can empty an entire treasury
- Cross-chain bridges concentrate enormous value in complex smart contract systems that connect otherwise isolated blockchains
- API-driven trading means that exchange credentials, if compromised, allow automated theft at machine speed
The industry already lost a record 3.3 billion dollars in 2025. Q1 2025 alone saw 1.64 billion dollars stolen across 40 incidents, the worst quarter in crypto history and a 4.7x increase over the same period in 2024.
The Bybit hack of Feb. 21, 2025, stands as the largest single crypto theft ever at approximately 1.5 billion dollars. The FBI attributed it to North Korea's TraderTraitor group, also known as Lazarus Group. Attackers compromised a Safe{Wallet} developer's workstation through social engineering, stole AWS session tokens, bypassed multi-factor authentication and injected malicious JavaScript into the wallet frontend. When Bybit employees approved what appeared to be a routine transfer, the altered interface redirected funds to attacker-controlled addresses.
That attack was a supply chain compromise, not a smart contract exploit.
It highlights a pattern now visible across the industry. Access control failures caused 53 percent of 2025 losses at 2.12 billion dollars. Smart contract bugs accounted for just 12.8 percent.
These are precisely the types of multi-step attack chains that Mythos has demonstrated it can plan and execute.
The UK AI Security Institute confirmed in an independent evaluation on Apr. 13 that Mythos succeeded 73 percent of the time on expert-level cybersecurity challenges that no model could complete before April 2025.
Also Read: Crypto Funds Pull $1.1B In Best Week Since January As Risk Appetite Returns
Exchanges, Custodians and Trading Desks as the First Pressure Point
Centralized exchanges and custodians sit at the intersection of massive capital pools and complex operational systems. They manage hot wallets, cold storage infrastructure, internal transfer approval workflows, employee access controls and API connections to thousands of trading bots. Each layer represents a potential attack surface.
The Bybit incident demonstrated how a single compromised developer workstation, combined with social engineering and frontend manipulation, could bypass even multi-signature cold wallet protections.
North Korean hackers alone stole 2.02 billion dollars from the crypto industry in 2025, up from 1.34 billion in 2024.
Key management is the central vulnerability. Private keys must be stored, transferred and used in signing ceremonies. Each step introduces human and software dependencies. A model like Mythos, which can autonomously chain three to five vulnerabilities into a sophisticated end-to-end exploit, could target these operational seams far more efficiently than human attackers.
Institutional trading desks face their own exposure. Most rely on API connections to multiple exchanges, with credentials stored in cloud infrastructure. An AI-powered attacker that compromises a trading firm's API keys could execute unauthorized trades, withdraw funds or manipulate order books, all at speeds that outpace human detection.
In a Linux kernel test documented in the System Card, Mythos autonomously discovered and chained multiple vulnerabilities to escalate from an ordinary user to complete machine control.
Nicholas Carlini, an Anthropic research scientist, noted the model can chain together three, four or sometimes five vulnerabilities in sequence.
Also Read: Bittensor's Most Powerful Builder Just Quit And Called The Whole Thing A Lie
What It Could Mean for DeFi and Web3 Infrastructure
Beyond centralized exchanges, the decentralized finance ecosystem presents an even more fragmented attack surface. DeFi protocols rely on interlocking layers of smart contracts, oracle feeds, governance mechanisms and cross-chain bridges.
Cross-chain bridges have historically been among the most vulnerable components.
The Wormhole exploit of February 2022 drained 326 million dollars after attackers bypassed signature verification through a deprecated function. The Ronin bridge hack of March 2022 saw 615 million dollars stolen after five of nine validator keys were compromised through social engineering.
Oracle dependencies add another dimension of risk.
DeFi protocols rely on price feeds from external data sources to execute lending, liquidation and trading functions. A compromised oracle can trigger cascading liquidations across multiple protocols simultaneously.
Smart contract audits, while necessary, are demonstrably insufficient. Data from AnChain.AI showed that 91.96 percent of hacked smart contracts had been audited, with some reviewed multiple times by reputable firms. The Cetus Protocol exploit of 2025 drained roughly 220 million dollars through a rounding bug in a third-party math library on the Sui blockchain, a type of subtle vulnerability that traditional audits often miss.
Mythos scored 100 percent on the Cybench benchmark and 83.1 percent on CyberGym, compared to Claude Opus 4.6's 66.6 percent.
On the OSS-Fuzz corpus, it achieved full control flow hijack on 10 separate, fully patched targets at the highest severity level. These capabilities, if replicated in open-weight models, would allow attackers to probe smart contract ecosystems with a level of thoroughness and speed that manual auditing cannot match.
Also Read: Brian Armstrong Backs CLARITY Act After Rejecting It Twice — What Changed
Why This Matters for Traders, Not Just Security Teams
Cybersecurity events in crypto do not stay confined to security teams. They become price events, liquidity events and volatility events. The Bybit hack triggered immediate market-wide selling pressure as traders rushed to assess counterparty exposure.
Bitcoin (BTC) and Ether (ETH) both dropped sharply in the hours following the Bybit disclosure.
Tokens associated with affected protocols tend to show a median 61 percent decline over six months after a major hack. Nearly 80 percent of crypto projects that suffer large-scale exploits never fully recover.
For traders, the implications of Mythos-class AI capability are concrete:
- A zero-day exploit targeting a major exchange could trigger a flash crash across correlated assets before most participants understand what happened
- A bridge exploit could fragment liquidity across chains, widening spreads and creating arbitrage dislocations
- A compromised oracle feed could cause cascading liquidations in lending protocols, amplifying downside volatility
- API key theft from a major market maker could distort order books and create false signals for algorithmic traders
Approximately 65 percent of crypto trading volume is now executed by AI-driven systems.
These systems rely on API connections, exchange infrastructure and on-chain data integrity. Any compromise of those inputs directly affects automated trading outcomes.
The speed dimension is critical. CrowdStrike's 2026 Global Threat Report showed that average eCrime breakout time has collapsed to 29 minutes. In a 24/7 market with no circuit breakers, that is more than enough time to drain a protocol and dump stolen tokens before most security teams can respond.
Also Read: Santiment Data Shows XRP Pessimism At Levels That Preceded Past Rallies
Could Mythos Accelerate Consolidation in Crypto?
The crypto industry is already bifurcating along security lines. Major exchanges are investing aggressively in infrastructure and compliance. Smaller protocols are struggling to keep pace.
Coinbase completed its 2.9 billion dollar acquisition of Deribit in December 2025, the largest crypto acquisition in history. The deal expanded institutional trading and custody capabilities. Binance achieved a 96 percent reduction in direct exposure to illicit funds between 2023 and 2025. Crypto.com secured 120 million dollars in digital asset insurance through Aon and Lloyd's of London.
Insurance premiums for crypto custodians rose 25 to 30 percent after Q1 2025 attacks. Smaller firms face steeper increases or outright denial of coverage.
The global crypto security market is projected to grow at roughly 14 percent annually through 2036 as spending on defenses accelerates.
The talent bottleneck compounds the problem. A global shortage of engineers specializing in zero-knowledge proofs, multi-party computation and post-quantum cryptography constrains security development across the industry. Larger firms can attract this talent with higher compensation. Smaller protocols cannot.
Halborn's report on the top 100 DeFi hacks from 2014 to 2024 found total losses of 10.77 billion dollars. Only 19 percent of hacked protocols used multi-signature wallets. Just 2.4 percent used cold storage.
If Mythos-class AI models proliferate, the security gap between well-funded institutions and smaller protocols will widen further. Exchanges that can afford AI-powered red-teaming, continuous vulnerability scanning and real-time threat detection will harden their defenses. Protocols that cannot may face existential risk from a single exploit.
Also Read: TON Could Become 3.5x Cheaper Than Solana If Durov's Fee Cuts Go Through
The Bullish Counterargument
The same class of AI model that threatens crypto security could also strengthen it. Anthropic explicitly frames Project Glasswing as defense-first, and several of its partners are already applying Mythos to protective use cases.
AI-powered security tools are rapidly emerging across the crypto ecosystem.
CertiK combines expert manual review with AI and formal verification for smart contract audits, and is trusted by Binance, OKEx and Huobi. Nethermind AuditAgent uses AI-driven vulnerability detection and attack scenario simulation. Octane Security delivers continuous offensive intelligence for smart contracts. Chainalysis Hexagate provides adaptive real-time on-chain threat detection.
Binance's own risk controls prevented 6.69 billion dollars in potential losses for 5.4 million users in 2025.
That figure suggests AI-powered defense is already working at scale, even if it does not capture headlines the way exploits do.
The UK AI Security Institute added an important caveat to its evaluation of Mythos. AISI noted that its test ranges lack security features often present in real enterprise environments, such as active defenders and defensive tooling. Mythos may prove more effective at finding vulnerabilities in static codebases than at defeating actively monitored production systems.
Heidy Khlaaf of the AI Now Institute raised a methodological concern.
Anthropic did not compare Mythos against traditional static analysis tools or report false-positive rates.
Without that comparison, it is difficult to assess whether the model's vulnerability discovery is a net improvement over existing automated tools or primarily a more accessible interface. Forrester, however, published a detailed analysis concluding that Anthropic backed up its assertions with evidence, and called the capabilities legitimate rather than marketing.
Bruce Schneier, the renowned security researcher, called Project Glasswing a PR play by Anthropic but acknowledged the underlying capabilities are real. He warned that the world needs to prepare for an environment where zero-day exploits become abundant.
Also Read: Binance Launches Prediction Markets To Rival Polymarket In $20B Sector
Conclusion
Claude Mythos Preview represents a genuine inflection point in AI cyber capability. It is the first frontier model whose offensive security powers were deemed too dangerous for public release.
The cryptocurrency industry set a new record of 3.3 billion dollars in losses in 2025, and that was before AI-powered attacks became widespread.
The attack chains Mythos can autonomously execute, including multi-step vulnerability discovery, exploit development and operational compromise, map precisely onto the access control failures and supply chain attacks that already cause the majority of crypto losses.
The real question is not whether AI will transform crypto security.
It is whether Anthropic's six-month estimate before open-weight models achieve comparable capabilities proves accurate, and whether the industry can deploy AI-powered defenses fast enough to match the pace of AI-powered offense.