Flash News

Anthropic Mythos Model Vulnerability List Fed Still Denied Access

Anthropic's Claude Mythos Preview has discovered thousands of unknown system vulnerabilities and is only open to about 50 institutions through Project Glasswing, excluding the Federal Reserve.

The Federal Reserve had urgently convened bank CEOs to warn of risks, but as of mid-July, it is still seeking access, while firms like JPMorgan, Apple, and Google have already gained it.

Access restrictions are event-driven, with a divergence in the speed of critical infrastructure patching. Funding and security resources are directed towards prioritized institutions, benefiting early participants and putting pressure on systemically important financial institutions.

Source: Public Information

ABAB AI Insight

Anthropic launched Mythos Preview in April 2026, which far exceeded expectations in autonomously discovering zero-day vulnerabilities in operating systems and browsers. The company chose not to publicly release it and initiated the Glasswing program to give defenders a first-mover advantage.

In terms of capital allocation, Anthropic prioritized access for tech giants and large banks, using resources for controlled patch collaboration, motivated by balancing the risk of model capability leakage with global critical software security.

Similar cases can be seen where government agencies often lag behind the private sector in early vulnerability bounty and responsible disclosure programs. Currently, we are in an expansion phase where AI-driven vulnerability discovery is spreading from laboratories to real critical systems.

This essentially represents a technological substitution, with the mechanism being that cutting-edge models significantly compress the vulnerability discovery cycle, allowing private companies to effectively control the release order of global software security information.

ABAB News · Cognitive Law

  1. The discoverer of the vulnerability decides who knows first.
  2. Systemically important institutions may not necessarily receive defensive tools first.
  3. The AI capability window burns on a monthly basis.

Source

·ABAB News
·
2 min read
·8 hrs ago
分享: