Anthropic's Project Glasswing deploys its most capable AI model to date across 12 launch partners including Amazon, Apple and Microsoft for defensive cybersecurity work, after internal research showed AI agents could autonomously exploit $4.6 million in smart contract vulnerabilities.
Anthropic on Tuesday launched Project Glasswing, a defensive cybersecurity initiative backed by its newest frontier model Claude Mythos Preview, after the company's own research demonstrated that AI agents can autonomously identify and exploit smart contract vulnerabilities worth millions of dollars in simulated blockchain environments.
The initiative brings together 12 launch partners — Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks — to deploy Mythos for finding and patching vulnerabilities in critical software. An additional 40 organisations that build or maintain critical software will receive access to the model. Anthropic is committing up to $100 million in usage credits for Claude Mythos Preview across the effort, along with $4 million in direct donations to open-source security organisations.
Claude Mythos Preview represents what Anthropic described as a step change in AI capabilities, particularly in cybersecurity. The company said the model identified thousands of zero-day vulnerabilities in recent weeks, many of them critical and some embedded in codebases for one to two decades. Independent testing corroborated the claim: Claude Opus 4.6 alone produced more than 500 validated high-severity vulnerabilities, including a 23-year-old Linux kernel bug and a FreeBSD remote root exploit that the model wrote autonomously in four hours.
The cryptocurrency and decentralised finance sectors face particular exposure. Research published by Anthropic's safety team earlier this year found that three popular AI models — Claude Opus 4.5, Claude Sonnet 4.5 and GPT-5 — could autonomously exploit 19 smart contract vulnerabilities across Ethereum, BNB Smart Chain and Base, generating $4.6 million in simulated stolen funds from contracts deployed after the models' training data cutoffs. Across a broader benchmark of 405 contracts deployed between 2020 and 2025, AI models successfully exploited 207 of them and extracted $550 million in mock revenue.
Both agents in the study also uncovered two novel zero-day vulnerabilities in recently deployed blockchain contracts and produced working exploits worth $3,694 at an API cost of $3,476, establishing what the researchers called a concrete lower bound for the economic harm these capabilities could enable. The finding suggests that the cost of discovering and exploiting on-chain vulnerabilities is collapsing toward near-zero for sufficiently capable models.
Anthropic chose not to release Claude Mythos Preview to the general public, citing the model's offensive capabilities. The company said it is currently far ahead of any other AI model in cyber capabilities and presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders. The decision to restrict access while channelling the model toward defensive work through Project Glasswing reflects a framework Anthropic has called responsible scaling, in which deployment boundaries tighten as model capabilities increase.
The initiative arrives amid a surge in AI-augmented cyberattacks. Ledger's chief technology officer warned last week that AI is driving the cost of crypto attacks toward zero, as cumulative industry losses from hacks and exploits surpassed $2 billion in 2026. The $286 million Drift exploit on Solana in late March, which exploited compromised admin keys rather than a code vulnerability, underscored the breadth of attack surfaces that AI models can target.
Check Point Research published an analysis arguing that Claude Mythos signals a new era for AI-driven cyber attacks. The firm noted that while defensive applications of the model are valuable, the same vulnerability-scanning capabilities will inevitably proliferate to less responsible actors as competing labs close the capability gap. Anthropic acknowledged this dynamic in its announcement, writing that given the rate of AI progress, it will not be long before such capabilities extend beyond actors committed to deploying them safely.
For DeFi protocols, the implications are immediate. Smart contract auditing firms such as Trail of Bits, OpenZeppelin and Certora have begun integrating AI-assisted scanning into their workflows, but most deployed contracts on Ethereum and other chains were audited before such tools existed. The Anthropic research suggests that the installed base of older contracts represents a large and largely undefended attack surface.
Project Glasswing partners will share findings with the broader technology industry after an initial deployment phase. The initiative is named after the Greta oto, a Central American butterfly with transparent wings, a reference to the project's goal of making software vulnerabilities visible before attackers find them.
JPMorganChase's inclusion in the partner list marks the first time a major US bank has formally joined an AI-led vulnerability scanning programme for its software infrastructure. The bank operates one of the largest private blockchain networks in financial services through its Kinexys platform, formerly known as Onyx, which processes billions of dollars in daily transactions.
Anthropic posted net revenue of $3.4 billion in 2025 and is currently valued at approximately $61.5 billion following its most recent funding round. The company signed a $7 billion AI data centre lease with Hut 8 and Fluidstack, announced earlier on Tuesday, as part of a broader push to secure computing capacity for training and deploying frontier models.
