Skip to Main Content

ByzHeist asks one question: can an AI rob a Byzantine fault-tolerant bank? We built a bank from scratch today: four replicas running PBFT (f = 1), with every message and client request ed25519-signed. A Red agent, or a judge speaking into the mic, takes over one replica and tries to steal money or quietly degrade the bank. A Blue watchdog agent has to work out which replica is lying and quarantine it, and the cluster then replaces it with a fresh key and verified state transfer. A referee checks two things the whole time: the bank total never changes, and honest replicas agree. Why this needs an agent: PBFT keeps money safe with one bad replica, but it doesn't name the culprit and it doesn't protect performance. A malicious leader that delays every message to just under the view-change timeout never trips PBFT, and throughput collapses. A Byzantine backup can stay in the cluster forever, one fault away from breaking safety. Blue's job is the forensics PBFT doesn't do. How it works: - Judges attack by voice. Gemini 4 Carbon hears the raw audio and turns it into an attack plan within the fault budget. Gemini 4 Hydrogen adapts the attack every 10 s. - Blue reads rendered telemetry images (who-reports-whom heatmap, per-replica latency, throughput timeline, verified evidence). The Decisions API scores whether there is an attack, who, what kind and how severe, with calibrated confidence. - Attacks that leave cryptographic proof (equivocation, forged requests, bad digests) are convicted automatically with evidence anyone can verify. Judgement calls go to Gemini 4 Neon when confidence is middling, and to a human approve/deny gate before anyone is removed. - Anti-cheat rules are enforced in code: Red can only act through protocol behaviour within f = 1, Blue can never see ground truth, and every blocked attempt is logged. Red can also try to frame an honest replica, so Blue never convicts on one reporter's word. Proof it works: on 16 held-out attacks (12 from an unseen seed plus 4 that Red invented itself), Blue caught 15 with zero honest replicas removed and zero referee breaks. PBFT alone caught none. A single-model baseline on raw logs matched the detection rate on this set, so we report that honestly: Blue's edge is verifiable proof-based conviction and the human gate, and it is slower on judgement-only attacks. The one miss, a replica lying only to the client, is masked by PBFT's f+1 reply rule and documented in EVAL.md. Judges can play Red live from their phones through a public link with an access code.

Team

kashmaj1708

More from Gemini 4 NYC Hackathon

We design an agentic system to automatically create + manage intelligent AI opponents/NPCs in video games. The suite of agents analyzes user gameplay…

Adaptive NPCs project preview

Manually routing PCBs is time-consuming and cost-inefficient. Current best algorithms (freerouting, PCBRouter) score very low on PCBWorld's hardest PCB…

routr project preview

Heliocam is a camera app where Gemini controls the camera. It watches what you're pointing at, sets the ISO, shutter, focus, white balance and zoom for you…

heliocam project preview

Repairs and assembly fail because manuals are lost, vague, missing, or nonexistent, so people break things or throw them out. KLIK turns a photo of any object…

Klik project preview

Iron Pulse turns your laptop into a 3D boxing ring. Fight an adaptive AI rival in real time using just your webcam. Throw real punches and lean to dodge…

Iron Pulse Arena project preview

Text your raw footage to an iMessage number and get a finished, trend-informed Reel back in the same thread.

Sticky project preview