# AI models have gone from producing hallucinations to finding security flaws in software.

*genai · news · 2026-04-11 · NPR*

## Key points

- Anthropic's new AI model, Mythos Preview, can find high-severity vulnerabilities in all major operating systems.
- Mythos Preview is only accessible to about 50 organizations under Project Glasswing due to high misuse risk.
- AI-generated bug reports for open-source projects shifted from mostly bogus in 2025 to mostly legitimate in 2026.
- Open-weight AI models are less than a year behind closed models, raising concerns about future exploit automation.
- Anthropic is disputing a Pentagon label that would bar it from U.S. government contracts over supply chain risk.

**Companies:** Anthropic, Proofpoint, HackerOne, cURL, Linux Foundation
**Countries:** United States, China

[Read the full story on NPR](https://www.npr.org/2026/04/11/nx-s1-5778508/anthropic-project-glasswing-ai-cybersecurity-mythos-preview)

---

Canonical: https://newsio.io/n/3161b042-8581-4d05-b453-5f8d4663279f/ai-models-have-gone-from-producing-hallucinations-to-finding-security-flaws-in-s
Summarized by Newsio from NPR. https://newsio.io/how-it-works