The 2026 tracks: four themes, one build day, everyone demoed.
New to hackathons? We've got you covered.
Attack AI Systems
[CLASSIFICATION: RED] OFFENSIVE PROTOCOL"Break the chain of trust."
Exploit the trust autonomous agents run on: decision systems, agent-to-agent links, self-governing AI.
Example projects: Prompt injection toolkit, agent jailbreak framework, AI-powered phishing detector evasion
Defend AI Systems
[CLASSIFICATION: BLUE] DEFENSIVE PROTOCOL"Guard the protocol."
Detect, harden and self-heal faster than attackers pivot.
Example projects: Real-time prompt injection detector, agent behavior monitor, automated security scanner for LLM apps
Verify Without Exposing
[CLASSIFICATION: CIPHER] TRUST PROTOCOL"Trust without compromise."
Prove integrity without exposing secrets: privacy-preserving verification and trust anchors.
Example projects: ZK proof for AI model authenticity, privacy-preserving credential verification, encrypted inference pipeline
Build Something New
[CLASSIFICATION: UNBOUND] UNDEFINED PROTOCOL"Write your own protocol."
Novel attacks, experimental defenses, or something nobody's thought of yet. You define it.
Example projects: AI-assisted threat hunting, autonomous red team agent, security copilot for code review
Multiple ways to participate. All paths lead to demos.
Start from zero at kickoff. Ship by sundown.
Bring something you've built. Separate judging.
Arrive alone, leave with a team.
How judges scored. Weights differed for BUILD_NEW and SHOWCASE.
A novel take on AI × Security.
Does it work, and how hard was it?
Could it ship? Does it solve a real problem?
Clear, working, under 3 minutes, handles Q&A.
Weighted toward innovation and what you built from scratch.
Weighted toward technical depth and real-world impact.
2-4 operators per team. Solo hackers welcome.
3 minutes to present. 2 minutes Q&A. No fluff.
100 operators maximum.
March 2027 · San Francisco. New tracks, same rule: everyone demos.
Get Notified