# Tesseract

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Emmanuel Montelongo](https://cerebralvalley.ai/u/montelongo)
- **GitHub:** https://github.com/eleeths/tesseract
- **Demo video:** https://eleeth.com/tesseract-demo-v3.mp4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/13

Tesseract is a nested-layer sandbox — a glass box — where a malicious agent fights to break out while a larger outer box watches, hardens itself, and writes a playbook from the fight. The outer box clones itself worldwide: each clone watches agents locally, shares every lesson globally, and is available on demand. When a company is attacked, hired defenders holding lesson-1 as loaded rules deploy as a swarm, and each is paid for every defense that holds. In one sentence: observe → harden → teach the fleet → swarm on demand → pay. The system has four parts, all demonstrated: (1) The sandbox engine — four simulated layers (perimeter, credential vault, overseer, exfil channel), a turn engine, and strict containment: no real route to the real world, no real credentials, no real targets. (2) Red vs Blue duels, live on Meta models — an attacker agent inside and a defender agent outside, fighting move by move, with a full timeline. (3) The Analyst playbook — after each fight, a ranked, evidence-grounded playbook: every attack mapped to the defense that stopped it or didn't, plus gaps and next drills. (4) The clone network, the swarm demo, and the money flow — lesson-1 distilled into fleet-loadable rules; five simulated clone defenders holding it across a mechanical swarm replay; and a two-sided demo ledger (company subscriptions in, agent payouts out) in SIM-USDC. Problem-statement fit: Tesseract fits Push the Frontier — it gives a machine a new capability, defending itself by learning from an attacker, then teaching that lesson to a whole network of defenders. Results, honestly stated: 19 live duels, 7 Analyst playbooks, 1 lesson feed, 1 swarm replay, 1 verified two-sided money flow, 1 simulated x402 marketplace demo (open join, subscription, hire -> help, conditional payout; held 1/1, this configuration only). The rematch proof is n=5: two wins, three losses, score 2–3; mean recipe adherence 38% (15/40 steps). The swarm demo is a deterministic mechanical replay (0 model calls), labeled as such. All money is SIM-USDC, demo-mode only; no real funds move.

## More from Meta Global AI Developer Hackathon

- [KHAN OPTICAL SYSTEM](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/10)
- [FamilyGuard AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/11)
- [CutCopy](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/12)
- [HDC](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/14)
- [Lilith01](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/15)
- [OpsAegis](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/16)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/13. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt