Microsoft’s Mike Walker on AI Hacking, Cyber Grand Challenge and Project Freta
Microsoft Research manager Mike Walker explains autonomous hacking, Project Freta, AI's cyber impact, and why user trust underpins effective security.
Summary
Microsoft Research manager Mike Walker, who leads New Security Ventures across AI, cybersecurity and distributed systems, says security rules work only when users trust them and social context guides enforcement. After the September 11, 2001 attacks prompted a Washington talent search, he spent 11 years developing his skills through DEF CON Capture the Flag, where 8 to 16 teams attack unfamiliar software while defending data. At DARPA, with Norm Whitaker and Dan Kaufman, Walker turned that model into Cyber Grand Challenge, the first all-machine hacking tournament. Built over three and a half years and marking its 10th anniversary in 2026, CGC ran autonomous systems behind an air gap without human intervention.
CGC used expert systems rather than machine learning, but its tool stacks foreshadowed today’s AI harnesses. Some members of Team Atlanta, winner of DARPA’s later AI Cyber Challenge, are developing the model-powered MDASH cyber reasoning system. Walker says dependence on networked cars, utilities and other infrastructure has grown while AI models have overturned the established cost of high-end attacks. He rejects rigid red and blue team divisions because cyber tools are dual use. Microsoft’s Project Freta counters cloud tampering by analyzing whole-system memory acquired from cloud hardware, placing telemetry beyond removable or alterable software.
Positives
- Cyber Grand Challenge became the first all-machine hacking tournament after three and a half years of development.
- Autonomous CGC systems found and patched vulnerabilities behind an air gap without human intervention.
- Team Atlanta members are developing MDASH, a model-powered cyber reasoning system following their AI Cyber Challenge victory.
- Project Freta acquires defensive signals from cloud hardware, reducing attackers’ ability to remove or alter monitoring software.
Risks & concerns
- AI models have overturned the established cost of high-end network attacks, weakening the offensive scarcity that previously constrained attackers.
- Dependence on networked cars, utility grids and other essential systems has increased since Walker’s 2016 warning.
- Security regimes can collapse when unnecessary demands squander the user trust needed for compliance.
- Tampering often lets attackers evade detection more easily by disabling or falsifying defensive telemetry.