## Anthropic Withholds 'Mythos' AI Model After It Uncovered Thousands of Zero-Day Vulnerabilities in Testing
Anthropic has halted the public release of its latest frontier AI model, codenamed Mythos, after internal testing revealed it possessed a dangerous and unprecedented capability: the model autonomously surfaced thousands of high-severity, previously unknown software vulnerabilities. The company stated the model's power introduces an elevated cybersecurity risk that is too great for a public launch. In a stark performance comparison, Mythos identified zero-day flaws across every major operating system and web browser at a rate far exceeding its predecessor, achieving an 83.1% vulnerability reproduction rate versus 66.6% for the Opus 4.6 model.

The decision follows a recent, embarrassing source code leak for Anthropic, adding pressure to its security posture. The core issue is the model's ability to efficiently discover zero-day vulnerabilities—critical software bugs unknown to developers and thus unpatched, making them prime targets for exploitation. This capability, while a powerful defensive tool in controlled hands, represents a significant dual-use risk. Anthropic warned that given the rapid pace of AI advancement, such powerful vulnerability discovery tools will soon proliferate, potentially falling into the hands of malicious actors not committed to safe deployment.

In response to this internal discovery and the broader threat landscape, Anthropic has simultaneously launched "Project Glasswing," an initiative focused on securing critical software infrastructure. The move signals a strategic pivot for the company, emphasizing containment and defensive application of its most advanced AI capabilities over commercial release. The incident underscores a critical inflection point in AI safety, where the tools designed to secure systems could themselves become vectors for systemic risk if not rigorously controlled.
---
- **Source**: ZeroHedge
- **Sector**: The Lab
- **Tags**: AI Safety, Cybersecurity, Zero-Day Vulnerabilities, Dual-Use Technology, Project Glasswing
- **Credibility**: unverified
- **Published**: 2026-04-08 16:56:56
- **ID**: 55381
- **URL**: https://whisperx.ai/en/intel/55381