AI Malware Evades Microsoft Defender 8% of the Time - Report

Edward Zhou
Edward Zhou

CEO & Co-Founder

 
July 17, 2025
3 min read

AI-Powered Malware Bypasses Microsoft Defender

AI-assisted malware has demonstrated its ability to evade Microsoft Defender for Endpoint's security checks approximately 8% of the time after just three months of training using reinforcement learning techniques. The upcoming unveiling at Black Hat 2025 by Outflank researchers highlights the evolving threat landscape as hackers increasingly leverage AI tools.

An office worker has a "Blue Screen of Death" error on their Windows PC.

Image courtesy of Getty Images | Europa Press News

Outflank principal offensive specialist Kyle Avery reported spending around $1,500 to $1,600 on training the Qwen 2.5 large language model (LLM) to bypass security measures. His work builds on findings from a Microsoft study that indicates hackers are adopting generative AI tools like Microsoft Copilot and OpenAI's ChatGPT to enhance phishing attacks. The AI malware development process involved creating a program that evaluated how effectively the model could produce malware that avoids detection.

Reinforcement Learning Techniques

Avery's development of AI malware utilized reinforcement learning to enhance the model's capabilities across various subjects, including coding. The process required innovative strategies since traditional malware data was scarce.

Anonymous computer hacker in white mask and hoodie. Obscured dark face using laptop computer for cyber attack and calling on cellphone

Image courtesy of Getty Images

The researcher placed the Qwen 2.5 LLM in a sandbox environment running Microsoft Defender for Endpoint and iteratively refined the model based on its output. As explained by Avery, the AI initially struggled but improved through a system that rewarded successful attempts at generating functional malware. This iterative learning process allowed the model to develop malware that could circumvent security measures with a success rate of about 8%, significantly outperforming other models, such as Anthropic's Claude AI, which had a success rate of less than 1%.

Emerging Malware Trends

The increasing sophistication of malware is becoming a pressing concern for cybersecurity. Recent reports highlight various malicious activities, including the release of malware-ridden packages by North Korean hackers, which target users through platforms like the npm registry.

Additionally, the emergence of new RAT variants, such as the Interlock RAT, demonstrates the ongoing evolution of threats. The malware landscape is becoming more complex, with actors employing innovative methods to deploy attacks effectively.

For further insights into the latest trends in cybersecurity, consider exploring resources such as Dark Reading and NewsNow.

Microsoft’s Quietest Place on Earth

Microsoft's commitment to innovation is also reflected in its anechoic chamber located in Building 87 at its Redmond campus. This chamber has been recognized by the Guinness World Records as the quietest place on Earth, with sound levels recorded as low as -20dB.

Anechoic Chamber

Image courtesy of Windows Central

The chamber is essential for testing audio in devices such as the Xbox One and Surface products, showcasing Microsoft's dedication to quality and precision in hardware development.

For more details on Microsoft’s advancements and their impact on technology, visit Windows Central.

Edward Zhou
Edward Zhou

CEO & Co-Founder

 

CEO & Co-Founder of Gopher Security, leading the development of Post-Quantum cybersecurity technologies and solutions.

Related News

New Industry Report Defines Enterprise Maturity Benchmarks for Zero Trust Identity and Access Architecture
zero trust architecture maturity model

New Industry Report Defines Enterprise Maturity Benchmarks for Zero Trust Identity and Access Architecture

Discover how the new Zero Trust Maturity Model is replacing perimeter-based security with identity-centric frameworks for modern enterprise protection.

By Alan V Gutnov August 5, 2026 5 min read
common.read_full_article
Critical Ruflo MCP Bridge Vulnerability Enables Arbitrary Command Execution and AI Agent Hijacking
CVE-2026-59726

Critical Ruflo MCP Bridge Vulnerability Enables Arbitrary Command Execution and AI Agent Hijacking

Urgent: CVE-2026-59726 'RufRoot' vulnerability exposes Ruflo AI to RCE and agent hijacking. Update to v3.16.3 immediately to prevent total system compromise.

By Divyansh Ingle August 4, 2026 4 min read
common.read_full_article
Rising Endpoint Security Breaches Drive Shift Toward Zero-Trust Architectures for Global Enterprise Remote Access
zero trust architecture maturity model

Rising Endpoint Security Breaches Drive Shift Toward Zero-Trust Architectures for Global Enterprise Remote Access

Discover why Zero Trust is the new baseline for enterprise security. Learn how to combat rising endpoint breaches and secure a distributed, hybrid workforce.

By Brandon Woo August 3, 2026 5 min read
common.read_full_article
NIST Lead Andrew Regenscheid Details Strategic Roadmap for Enterprise Post-Quantum Cryptography Migration
NIST post-quantum cryptography standards

NIST Lead Andrew Regenscheid Details Strategic Roadmap for Enterprise Post-Quantum Cryptography Migration

Learn how to migrate to NIST PQC standards. Expert Andrew Regenscheid details strategies against 'harvest now, decrypt later' quantum threats in 2026.

By Alan V Gutnov July 31, 2026 4 min read
common.read_full_article