The SaaS-pocalypse Is Here: How Anthropic's AI Wiped $1 Trillion From Software Stocks—And Why the Worst May Not Be Over

Software stocks extended a brutal losing streak on Friday, April 10, with the iShares Extended Technology Software Sector ETF falling another 2.8%—bringing its three-day decline past 7% and its year-to-date loss to roughly 30%. Cloudflare dropped 11% on the day. ServiceNow shed 8.5%, closing near $83. Palo Alto Networks fell 7.2%, settling around $155. RingCentral, HubSpot, Datadog, CrowdStrike, Zscaler, and Palantir all posted significant losses. Analysts estimate the sector has shed $1–2 trillion in market value since January—what some are calling the largest non-recessionary drawdown in software in decades.

The Trigger: Anthropic's Claude Mythos and the Zero-Day Bombshell

The immediate catalyst was Anthropic's release earlier in the week of a powerful new AI model, reportedly called Claude Mythos, alongside related agentic infrastructure under the banner of Managed Agents. What rattled markets was not the model's benchmark scores—it was what the model did. In testing, Mythos autonomously identified and demonstrated exploits for zero-day vulnerabilities across every major operating system and browser, including a 27-year-old flaw in OpenBSD and a 16-year-old bug in FFmpeg—flaws that had survived decades of human security review.

Anthropic limited access to roughly 40 select technology organizations under a defensive program called Project Glasswing, citing the model's danger. That restraint, paradoxically, amplified investor fear. If a single AI system could shred the security assumptions baked into legacy codebases this efficiently, what did that imply about the durability of software infrastructure itself?

Analyst actions compounded the damage. UBS downgraded ServiceNow, citing autonomous AI agents as a structural threat to its workflow automation franchise. Citi downgraded multiple software names amid uncertainty around AI's disruption of legacy business models.

A Category Inversion, Not a Normal Cycle

Here is what the headline numbers obscure: this is not a valuation correction. It is a repricing of an entire logic.

Enterprise software was built to monetize friction—expensive builds, painful integrations, slow workflow changes, and the simple fact that users had to navigate applications to get work done. Each of those frictions is now under direct assault. Anthropic's own product roadmap makes the direction explicit: Claude Cowork is positioned not as a chat assistant but as a system that executes multi-step knowledge work autonomously. Managed Agents is designed for long-horizon tasks where orchestration, memory, sandboxing, and control increasingly belong to the AI layer—not the enterprise buyer.

The consequence is architectural. The application is being demoted from destination to callable backend. When that shift completes, seat-based pricing, module expansion, and UI engagement—the three pillars of SaaS revenue growth—lose their economic logic simultaneously.

What Is Actually Dying, and What Survives

The phrase "software is dead" is too blunt—and too convenient. What is dying is a specific form: the UI-heavy, seat-heavy, workflow-navigation-heavy product that earns rent by standing between users and outcomes.

The categories most exposed are horizontal workflow SaaS, light analytics layers, CRM extensions, support-ticket organizers, and low-differentiation productivity tools. These products exist primarily to help humans move through information. Agents absorb that job.

What survives—and likely consolidates value—is narrower: identity and authorization infrastructure, transaction rails, systems of record with legal or financial consequence, compliance engines, and runtime policy enforcement for agents themselves. Even cybersecurity bifurcates. Legacy alert-console vendors built around human analyst throughput face the same structural pressure as SaaS. The winners in security will be those governing machine-speed agent behavior—runtime isolation, exploit simulation, patch orchestration—not those counting human tickets.

The End-State Markets Are Beginning to Price

The most important insight is also the most uncomfortable for incumbents: "AI features" are not a moat. At best they are table stakes. At worst, they signal that management is defending the wrong architectural layer entirely.

The realistic end-state is not a world of universal agents floating above dead software. It is a compressed stack: a small number of powerful agent interfaces at the user layer, a lean set of governance and control-plane tools in the middle, and a durable but less visible substrate of systems of record and transaction engines beneath. The interface collapses; the trusted machinery persists.

The market is finally asking the right question: when agents become the default interface to work, which software companies still own the bottleneck? On April 10, 2026, the answer—for most of the sector—appears to be none.

not investment advice