AI Agents & Agentic Engineering
  • Context Engineering
  • Building AI Agents
  • Agentic Engineering Best Practices
  • Auto-Research and Self-Improving Agents
  • Self-Maintaining Agents
  • Scaling Long-Running Autonomous Coding
  • Data Agents
  • Agent Memory and Harness Ownership
LLM Internals & Training
  • Reinforcement Learning for LLMs
  • Cursor Composer 2
  • Training-Free GRPO
  • Prompt Repetition for Non-Reasoning LLMs
  • LLM Quantization
  • LLM Knowledge Bases
AI Safety & Security
  • Constitutional Classifiers
  • Project Glasswing
Graphics & Low-Level Programming
  • Shaders and GPU Programming
Personal Development & Mindset
  • Deep Focus and Productivity
  • Identity and Life Design
  • Self-Competition and Entrepreneurship

Project Glasswing

Project Glasswing is an Anthropic-led initiative uniting AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks to use frontier AI for defensive cybersecurity. The goal: find and fix vulnerabilities in critical software before attackers can exploit them.

Source: Anthropic

Claude Mythos Preview

The initiative centers on Claude Mythos Preview, an unreleased frontier model that Anthropic says can "surpass all but the most skilled humans at finding and exploiting software vulnerabilities."

Capabilities

Mythos Preview has autonomously found thousands of high-severity zero-day vulnerabilities in every major operating system and every major web browser. Notable examples:

  • A 27-year-old vulnerability in OpenBSD (one of the most security-hardened OSes) allowing remote crash of any machine by connecting to it
  • A 16-year-old vulnerability in FFmpeg in a line of code automated tools had hit 5 million times without catching
  • Autonomously found and chained together multiple Linux kernel vulnerabilities to escalate from user to root

Most of these were found entirely autonomously, without human steering.

Benchmark Performance

BenchmarkMythos PreviewOpus 4.6
SWE-bench Verified93.9%80.8%
SWE-bench Pro77.8%53.4%
SWE-bench Multilingual87.3%77.8%
Terminal-Bench 2.082.0%65.4%
CyberGym83.1%66.6%
GPQA Diamond94.6%91.3%
BrowseComp86.9%83.7%
OSWorld-Verified79.6%72.7%

The Cybersecurity Argument

AI models have reached a level where the cost, effort, and expertise required to find and exploit vulnerabilities have dropped dramatically. The window between discovery and exploitation has collapsed from months to minutes. Without defensive action, these capabilities will proliferate to malicious actors.

The same capabilities that make AI dangerous for offense make it invaluable for defense. Project Glasswing aims to give defenders a durable advantage.

Partner Commitments

Partners receive access to Mythos Preview for:

  • Local vulnerability detection
  • Black box testing of binaries
  • Securing endpoints
  • Penetration testing of systems

Financial Commitments

  • $100M in usage credits for Mythos Preview across participants
  • $2.5M to Alpha-Omega and OpenSSF through the Linux Foundation
  • $1.5M to the Apache Software Foundation
  • Model pricing after research preview: $25/$125 per million input/output tokens
  • Available via Claude API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry

Availability

Mythos Preview is not planned for general availability. Anthropic's goal is to develop safeguards (including constitutional-classifiers) that can detect and block dangerous outputs before deploying Mythos-class models at scale. New safeguards will launch first with an upcoming Claude Opus model.

90-Day Public Report

Within 90 days, Anthropic will report publicly on:

  • Vulnerabilities fixed and improvements made
  • Practical recommendations for evolving security practices
  • Vulnerability disclosure, patching automation, software development lifecycle, supply-chain security, standards for regulated industries

Related Topics

  • constitutional-classifiers - Safeguard technology required before Mythos-class deployment
  • scaling-autonomous-coding - Autonomous agents applied to security rather than development