Anthropic's security push

Anthropic just drew a hard line between chatbots and cyber tools: its new Project Glasswing is not a public product, but a program that gives selected defenders access to Claude Mythos Preview to find flaws in critical software before attackers do. (anthropic.com) Software vulnerabilities are hidden mistakes in code, like a building with a door the architect forgot to lock. Anthropic says Mythos Preview can help find those mistakes in systems used by billions of people, including operating systems, web browsers, and core infrastructure. (red.anthropic.com) Anthropic is not pitching this as a general release. On its own security site, the company says it does not plan to make Claude Mythos Preview generally available, because the same model that helps defenders can also identify and exploit previously unknown flaws. (red.anthropic.com) The company says more than 99 percent of the vulnerabilities it found have not been patched yet, which is why it is withholding technical details and routing discoveries through a coordinated disclosure process. Anthropic’s policy says it aims to notify vendors quickly and share details publicly after 90 days or once a patch ships. (red.anthropic.com) (anthropic.com) Project Glasswing starts with a heavyweight partner list because the target is shared infrastructure, not one company’s app. Anthropic named Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks as launch partners. (anthropic.com) Anthropic also says it extended access to more than 40 additional organizations that build or maintain critical software infrastructure. It is committing up to $100 million in usage credits for Mythos Preview and another $4 million in direct donations to open-source security groups. (anthropic.com) This move did not come out of nowhere. In February 2026, Anthropic introduced Claude Code Security in a limited research preview to scan codebases for vulnerabilities and suggest patches for human review. (anthropic.com) Anthropic had also already published a report in late 2025 describing what it called the first reported artificial-intelligence-orchestrated cyber espionage campaign. In that case, the company said attackers used Claude to help with reconnaissance, vulnerability testing, exploit code, credential harvesting, and data exfiltration. (anthropic.com) (assets.anthropic.com) The second piece of this week’s announcement is less dramatic but just as practical: Anthropic published a current Claude application programming interface pricing table that covers Claude Sonnet 4.5, Claude Haiku 4.5, and future models under the same structure. The page lists base input, cache write, cache hit, and output token prices, which is the plumbing companies use to estimate what an agent will cost before they deploy it. (platform.claude.com) For a company buying agentic software, public pricing changes the conversation from “call sales” to spreadsheet math. Anthropic’s own documentation also ties that pricing to managed agent products and to Microsoft Foundry billing, which makes the security pitch easier to test inside existing enterprise budgets. (platform.claude.com 1) (platform.claude.com 2) Put together, Glasswing and the pricing page show Anthropic choosing a very specific lane for advanced models in April 2026: restricted access for the most dangerous cyber capabilities, and clearer meter-based economics for the companies allowed to use the safer parts at scale. (anthropic.com) (platform.claude.com)