Anthropic says its most powerful AI cyber model is too dangerous to release publicly — so it built Project Glasswing

Anthropic launched Project Glasswing pairing unreleased frontier model Claude Mythos Preview with 12 major partners and committed up to $100M in usage credits plus $4M in donations; the company also reported a >$30B annualized revenue run rate. Mythos Preview reportedly autonomously found thousands of high-severity zero-days (including an OpenBSD flaw and Linux kernel escalation) and will be priced post-preview at $25 per million input tokens and $125 per million output tokens, while Anthropic secured ~3.5 GW of compute via Google/Broadcom. Significant operational lapses (CMS leak, npm packaging error) and the logistical challenge of disclosing thousands of bugs create reputational and execution risks; the initiative is sector-moving and supports Anthropic's IPO narrative but warrants caution for partners and maintainers.

Analysis

This development accelerates a secular bifurcation: vendors that control telemetry, orchestration, and patch deployment will capture disproportionate value as automated vulnerability discovery scales. Expect cloud providers and hyperscale networking/infra suppliers to monetize detection and remediation workflows as a software-adjacent revenue stream, not just compute sell-through; that changes margin mix and upsells in existing enterprise contracts over 6–18 months.

Operational risk is the dominant near-term catalyst. A single high-profile mishandled disclosure or an exploit traced to an AI-originated chain could trigger regulatory intervention (export controls, mandatory disclosure rules) within months, compressing commercial addressable market and forcing more conservative productization cycles. Conversely, clean execution and demonstrable reductions in mean-time-to-patch will validate premium pricing for managed AI security services and shorten payback to 12–18 months for large customers.

The supply-chain winners are non-obvious: network ASIC and interconnect vendors that reduce latency and cost-per-inference will see durable demand even if model licensing commoditizes, while companies reliant on legacy on-prem security appliances will face structural margin pressure. Over a 2–3 year horizon, the most durable moats will belong to firms that combine proprietary telemetry, orchestration locks-in, and scalable managed services — not merely those who ship signature-based appliances.

Key Decisions for Investors

• Long CRWD (6–12 months): buy shares or a 9–12 month call spread sized as a 3–5% portfolio exposure. Thesis: leader in cloud-native endpoint telemetry with fastest path to monetize AI-driven detection; downside: 30–40% if competition compresses price or a macro drawdown hits tech multiples.
• Long AVGO (12 months): accumulate stock on 5–10% pullbacks. Thesis: benefits from higher datacenter networking and switch silicon demand as AI inference scales; reward: asymmetric 20–30% upside if hyperscale capex stays strong, risk: cyclical server market pullback trimming earnings by >15%.
• Long GOOGL or AMZN (6–12 months) vs Short CSCO (pair trade): equal notional long in a cloud vendor and short in Cisco. Thesis: clouds capture software/security upsell and recurring margins while legacy appliance vendors face slower transition; target 1.5:1 reward:risk over 6–12 months, close if regulatory headlines or macro shock flips cloud multiple performance.
• Buy NVDA 9–12 month calls (small allocation, 1–2% portfolio) as a convex bet on continued GPU scarcity and pricing power. Keep position small due to valuation risk; take profits on a 30–40% move and cut if implied volatility spikes >2x baseline or guidance misses.