Anthropic Mythos: The AI Too Dangerous to Release Is Now Opening to All
Anthropic Mythos: The AI Too Dangerous to Release, Now Opening to All
On May 28, 2026, Anthropic made a stunning announcement: the company's most powerful AI model — previously deemed "too dangerous" for public release — will be made available to all customers within weeks. The model, Claude Mythos, possesses unprecedented cybersecurity capabilities, including the ability to autonomously discover and exploit zero-day vulnerabilities across all major operating systems and browsers.
Anthropic had initially restricted Mythos to just 50 companies under "Project Glasswing," a defensive cybersecurity initiative. The decision to open it more broadly represents a dramatic shift, driven by competitive pressure and the realization that open-source models will catch up within 6-18 months.
The Numbers That Shocked the Industry
Mythos is not an incremental upgrade. It represents a "generational leap" in AI's ability to understand, analyze, and exploit software vulnerabilities.
| Benchmark | Mythos | Opus 4.6 | Improvement |
|---|---|---|---|
| CyberGym | 83.1% | 66.6% | +16.5% |
| SWE-bench Verified | 93.9% | ~70% | Highest on record |
| GPQA Diamond | 94.6% | ~80% | Highest on record |
| Firefox Exploit Generation | 181 successful exploits | 2 successful exploits | 90x improvement |
What Mythos Can Actually Do
- Found a 27-year-old vulnerability in OpenBSD's TCP stack — a system renowned for its security, broken by two specially crafted packets.
- Discovered a 16-year-old FFmpeg bug that survived 5 million fuzzing attempts.
- Identified thousands of zero-day vulnerabilities across Linux, Windows, macOS, Chrome, Safari, and Firefox, with over 99% unfixed at time of discovery.
- Autonomously chains 3-5 vulnerabilities into multi-step attacks, from initial access to full domain admin — something no previous AI could do.
Project Glasswing: Building the Shield Before the Sword
When Anthropic first discovered Mythos's capabilities, the company faced an unprecedented dilemma. The model could be used to defend critical infrastructure — or to destroy it.
Their solution was Project Glasswing, named after the glasswing butterfly whose transparent wings hide it from predators — a metaphor for the invisible vulnerabilities lurking in software.
| Category | Participants |
|---|---|
| Launch Partners | Amazon AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks |
| Key Infrastructure Orgs | 40+ organizations maintaining critical software infrastructure |
| Government Access | NSA, select US departments, UK financial regulators, Canadian regulators |
According to Anthropic, Mozilla reported that its latest Firefox release fixed 271 vulnerabilities identified by Mythos. The model has also identified more than 23,000 vulnerabilities across open-source projects, with an estimated 6,200 classified as high or critical severity.
Valuation Hits $965 Billion, Surpassing OpenAI
On the same day, Anthropic announced its valuation has reached $965 billion, surpassing OpenAI's estimated valuation for the first time. This represents a stunning reversal from just months ago, when OpenAI was widely considered the undisputed leader in frontier AI.
The valuation surge is driven by several factors: the unique defensive value of Mythos, enterprise demand for secure AI, and the perceived strategic advantage of Anthropic's cautious approach to deployment.
The Price of Power: $25/$125 per Million Tokens
Access to Mythos Preview, when it opens, will come at a premium price: $25 per million input tokens, $125 per million output tokens.
By comparison, Claude Opus 4.6 costs $5/$25 — making Mythos roughly 5 times more expensive. The price reflects not just the computational cost but the value of its unique cybersecurity capabilities. Anthropic has provided $100 million in credits to Project Glasswing partners, softening the impact for defenders while the ecosystem adapts.
Is It Worth It?
For organizations protecting critical infrastructure, the answer appears to be yes. A single undetected zero-day can cost millions. The ability to proactively find and fix vulnerabilities before attackers exploit them, at a fraction of the cost of traditional security audits, makes Mythos economically rational despite its high price tag.
Key Takeaways
- Anthropic will release Mythos-level AI to all customers within weeks — reversing its earlier "too dangerous to release" position.
- Mythos is 90x better at exploit generation than the previous best model, finding thousands of zero-day vulnerabilities.
- Project Glasswing partners include 12 founding members and 40+ critical infrastructure organizations, with government access from NSA and UK regulators.
- Anthropic's valuation has reached $965 billion, surpassing OpenAI for the first time.
- Pricing is $25/$125 per million tokens — 5x more expensive than Opus, but with unique defensive value.
- The clock is ticking — open-source models are expected to achieve similar capabilities within 6-18 months.
Sources & Methodology (as of May 31, 2026):
- NHK World — Anthropic wide release announcement
- 科技日报 (Science and Technology Daily) — Mythos cybersecurity analysis
- 至顶网 (ZDNet China) — Opus 4.8 and Mythos release timeline
- 绿盟科技技术博客 (NSFOCUS) — Technical capabilities and benchmarks
- eWEEK — Mythos public release prospects
- LLM Stats — Pricing and Project Glasswing details
- 电子工程专辑 (EET China) — Government and financial sector response
- 36氪 (36Kr) — Pricing and cost analysis
Stay Ahead of Tech Trends with Gzmato
Fast Chargers | USB-C Cables | Screen Protectors | Wireless Chargers | Power Banks | Travel Adapters
Special Offer: Use code TECH2026 for discount on your first order!
- Anthropic Mythos
- Claude Mythos
- AI cybersecurity
- zero-day vulnerabilities
- Project Glasswing
- Anthropic valuation
- AI safety
