# Blueshift

- **Event:** [Gemini 4 Argon SF Hackathon](https://cerebralvalley.ai/e/gemini-4-argon-sf)
- **When:** Sun, Oct 25 at 9:00 AM – 10:00 PM (PDT)
- **Where:** San Francisco, CA
- **Team:** [Ryan MacArthur](https://cerebralvalley.ai/u/tensor)
- **GitHub:** https://github.com/maceip/deepmind-argon
- **Demo video:** https://youtu.be/iiwZV3P4-Vc
- **Gallery:** https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/108

Blueshift is a frontier cyber intelligence platform that pairs 3 agents: red, green, and blue

Problem:
Defenders face a structural bottleneck: safety alignment locks them out of frontier red-team capabilities, and even when teams deploy uncensored offensive models, they cannot cleanly pipe raw exploit execution traces into defensive remediation agents. Consequently, static scanners flood teams with unverified theoretical findings, while standard coding agents break cross-file interfaces by patching code without full-repo context or live failure telemetry.

Solution:Fragments is an autonomous red-to-blue workbench powered by Gemini 4 Argon and Google Cloud Model Armor that chains live offensive exploitation directly into full-repository remediation:

1M-Context Recon (gemini-red): Ingests the target repository into Gemini 4 Argon’s 1M context cache to synthesize a tailored pentest suite on the fly, inspecting all bidirectional traffic through GCP Model Armor for prompt injections, jailbreaks, and Cloud DLP leaks.

Hardware-Gated Live Exploitation: gemini-red drives the target app inside an embedded 16:10 browser sandbox, pauses for native macOS Touch ID authorization, and executes a weaponized payload generated by remote abliterated red-team weights—proving the exploit on screen.

Chained Red-to-Blue Remediation (gemini-green): The live exploit trace feeds directly into gemini-green, which uses Gemini 4 Argon’s 1M continuous output window (thinking_level: high) to trace root causes across modules in one pass. It renders a coordinated multi-file code and Model Armor patch inline (Lines 95–99) inside a Unified/Split diff viewer.

Hot-Patch & Perimeter Sync (gemini-blue): Clicking [Apply Patch] hot-swaps the running server (v1 to v2), syncs perimeter rules to a GCP Maglev Load Balancer via gemini-blue, and automatically replays the browser exploit—verifying HTTP 403 Blocked with a 100% single-pass fix rate on cross-file vulnerabilities (vs. 33% on chunked baselines).

Built Today at the Hackathon (Oct 25, 2026):
Mounted on our open Compose Desktop window shell, 100% of the core architecture was built and verified today: the Gemini 4 Argon 1M client, the gemini-red / gemini-green / gemini-blue handoff pipeline, the on-the-fly pentest generator with GCP Model Armor and Cloud DLP telemetry cards, the live browser agent with macOS Touch ID gating, the multi-file Unified/Split diff viewer with version scrubbing (v1 to v2 to v3), the GCP Realtime SDK integration (Cloud Logging, Trace, Storage), and a 28-test end-to-end verification suite

## More from Gemini 4 Argon SF Hackathon

- [Argonauts](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/105)
- [Sightline](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/106)
- [MemoryWorld](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/107)
- [ExpLit](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/109)
- [PersonalStylist](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/110)
- [Argon Arcade](https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/111)

---

Markdown version of https://cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/108. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt