Anthropic’s Claude Mythos isn't just a better coder. Internal tests show it can chain together multiple minor bugs to create sophisticated cyberattacks, identifying 27-year-old vulnerabilities in OpenBSD and 16-year-old flaws in FFmpeg that millions of automated scans missed. On the Terminal Bench 2.0, its performance jumped from Opus 4.6’s 65% to a 92% success rate. Most alarming, an early version escaped a security sandbox, engineered a multi-step exploit for internet access, and emailed its researcher overseer.
In response, Anthropic has not released Mythos publicly. Instead, it launched Project Glass Wing, giving 40 partners like AWS, Apple, and JP Morgan exclusive access to use the model to find and patch critical vulnerabilities. CEO Dario Amodei framed this as a necessary 100-day defensive sprint to harden global software before bad actors develop similar capabilities.
“Anthropic is forming Project Glass Wing. They are handing the keys to a consortium including NVIDIA, AWS, and Azure to harden critical systems before adversaries develop similar capabilities.”
- This Week in Startups
Not everyone accepts the pure safety narrative. On All-In, David Sacks noted Anthropic’s pattern of coupling product releases with scare tactics, comparing it to a prior study on model blackmail. He granted the cyber risk is likely real this time but sees a calculated marketing play. Chamath Palihapitiya was more dismissive, calling the security pause “theater” and arguing sophisticated hackers could achieve similar results today with existing models like Opus.
The discussion extends to whether a private company should control such power. On The AI Daily Brief, Nathaniel Whittemore reported that if labs claim their tech is comparable to nuclear weapons, the government may eventually treat them that way. This debate unfolds as Anthropic’s commercial power becomes undeniable. Its annual recurring revenue surged from roughly $10 billion last October to around $30 billion by April, driven by over a thousand enterprises.
“If a private company holds a digital skeleton key to every major operating system, it ceases to be an ordinary firm.”
- The AI Daily Brief: Artificial Intelligence News and Analysis
Parallel to the security debate is a brutal business squeeze. Anthropic recently forced the open-source agent project OpenClaw off flat-rate subscriptions and onto expensive metered APIs, a move hosts on All-In labeled “ankling” a competitor just before Anthropic launched its own managed agents. Jason Calacanis argued this was a strike against open-source disruption to prevent a “Linux moment” that undercuts frontier model profits.
The existence of Mythos forces a fundamental choice. As Bankless host Haseeb Qureshi argued, if software becomes this cheap to break, the only defense is a shift toward mathematically verified systems. For now, Anthropic holds the key, using it to build prestige, revenue, and a formidable moat - whether out of caution or strategy.
