Huo Jingnan from NPR reports that AI models, once known for making errors, are improving at finding software security flaws. This can help developers fix problems faster, but it also raises concerns if hackers use these tools.
Anthropic, an AI research company, recently introduced a new model called Mythos Preview. The company says it could “reshape cybersecurity” by identifying serious vulnerabilities, including some found in major operating systems and web browsers.
The Anthropic’s Mythos Preview revealed the following:
- AI capabilities and potential benefits: Mythos Preview can detect and even exploit serious security weaknesses in software. This may help developers find and fix issues more effectively than before.
- Access and oversight: Anthropic is limiting access through Project Glasswing. About 50 selected companies and organizations can use the model. The Linux Foundation is one of the participants and is exploring how it could help improve open source software security.
- Risks and broader impact: Experts say the model could be misused to exploit systems, which is a concern for the cybersecurity community. At the same time, they note that most everyday users are not likely to face direct risk right now.
Below are additional highlights from Jingnan’s NPR article on how advances in AI are reshaping how software vulnerabilities are identified and managed, bringing both benefits and new challenges.
- Hackers are using AI to find bugs: Improved AI models in early 2026 led to a surge of false security reports for cURL, a widely used data transfer tool. Many of these reports were long and detailed, and likely created by AI. This made it harder for developers to spot real security issues.
- LLMs have now bypassed human capability for bug finding: The quality of bug reports for open source projects has improved, with more real issues being identified. This change is linked to Anthropic’s Opus 4.5 model, which appears to perform better than humans at finding bugs.
- Can AI find and fix security vulnerabilities bug patches? Daniel Stenberg, lead software developer behind cURL, who is not part of Project Glasswing, told NPR that AI is better at finding bugs than fixing them. Fixing issues often requires careful judgment and takes more time. He also points out that many important open source projects are not included in the initiative.
- Role of AI in Cyber security and cyber warfare: Advanced AI models from major labs include built-in safeguards, which help reduce misuse. However, more accessible open-weight models may soon reach similar capabilities and could be used to create harmful software. Some of these models are being developed by Chinese companies, which adds to concerns about global competition in AI. Alex Stamos, chief security officer at Corridor, an AI software security company, noted that by not releasing Mythos Preview to the public, Anthropic is giving developers and the U.S. more time to strengthen their defenses.
Acknowledgement: This summary was generated using AI writing tools as part of my AI Explorations process. For more information, please refer to the original articles listed below.