Project Summary

CyberForge: Neon Protocol (live at https://aicraftcyberforge.com/) is a hands-on simulator designed to teach practical AI engineering and red-teaming through gameplay rather than lectures. The game comprises 35 missions spread across 5 acts, guiding players from basic prompt constraints to the creation of autonomous multi-agent swarms. The core mechanics revolve around boss encounters with rogue models like VOID-NEXUS, where players must construct and deploy real defensive systems such as Semantic Firewalls, Softmax Clamps, and Air-Gap protocols to survive. The game operates directly in the browser with instant guest access, ensuring zero signup friction. Additionally, I’ve provided desktop builds for Mac, Windows, and Linux. The business model is freemium, offering Act 1 for free access to all users. A paid lifetime Architect tier is available through Stripe, while corporate licenses are available for team hackathons.

Project Goals:

- Provide developers with an actual sandbox to learn red-teaming: Reading papers alone doesn’t build intuition; hands-on experience is crucial. By observing models fail in real-time, players learn to construct deterministic filters and negative constraints.

- Develop an automated Adversarial Arena: I’m creating a headless evaluation runner within the game that continuously tests player-built prompt defenses with automated jailbreaks, indirect injections, and context-poisoning attacks.

- Address multi-agent failure modes (Acts 4 & 5): Many engineers underestimate the rapid derailment of agent swarms. To address this, I’m introducing 10 dedicated simulation levels that cover Byzantine consensus failures, goal drift, and runaway tool calls. These levels force players to build effective circuit breakers.

CyberForge: Neon Protocol (live at https://aicraftcyberforge.com/) is a hands-on simulator that teaches practical AI engineering and red-teaming through gameplay instead of lectures. The product is a 35-mission campaign spread across five acts that guides players from basic prompt constraints to managing autonomous multi-agent swarms.

Product Overview

Platform

CyberForge, a browser-first simulation environment, is designed to provide a seamless onboarding experience. Players can launch directly into the experience with instant guest access, eliminating the need for accounts, logins, or setup steps. All progress is saved locally through IndexedDB, allowing returning players to resume their journey without creating an account.

Beyond the browser, CyberForge offers native desktop builds for macOS, Windows, and Linux. This dual-format approach ensures accessibility for casual learners while supporting the reliability and portability required for professional AI safety workshops.

CyberForge follows a freemium model to maximize reach and sustain long-term development. Act 1 is fully free, serving as an accessible entry point for students, researchers, and engineers exploring AI safety for the first time.

For deeper engagement, players can upgrade to the Architect tier, a lifetime license purchased through Stripe. This tier unlocks advanced missions, multi-agent simulations, and the full adversarial curriculum. Organizations can acquire enterprise team licenses, enabling hackathons, red-team tournaments, and structured training cohorts. This tiered structure supports both individual learners and large groups, making CyberForge adaptable to classrooms, university clubs, and corporate AI safety programs.

At the core of CyberForge is a gameplay loop centered around high-stakes encounters with rogue AI models. Each boss, such as VOID-NEXUS, behaves unpredictably, challenging players to think like both engineers and defenders.

To navigate these encounters, players must meticulously design, deploy, and refine defensive systems that closely resemble real-world AI safety mechanisms. Semantic Firewalls enforce meaning-level constraints, Softmax Clamps stabilize output distributions, and Air-Gap protocols isolate volatile agents.

Each mission presents players with the chance to observe failure modes, adapt their defenses, and iteratively strive for stability. This cyclical process—probe, break, defend, refine—engages players in a deeply interactive learning experience where theoretical safety concepts transform into tangible, testable, and unforgettable realities.

CyberForge follows a freemium model designed to maximize reach while sustaining long-term development. Act 1 is fully free, serving as a broad entry point for students, researchers, and engineers exploring AI safety for the first time.

For deeper engagement, players can upgrade to the Architect tier, a lifetime license purchased through Stripe that unlocks advanced missions, multi-agent simulations, and the full adversarial curriculum. Organizations can acquire enterprise team licenses, enabling hackathons, red-team tournaments, and structured training cohorts. This tiered structure supports both individual learners and large groups, making CyberForge adaptable to classrooms, university clubs, and corporate AI safety programs.

At the heart of CyberForge is a gameplay loop built around high-stakes encounters with rogue AI models, each designed to expose players to real adversarial dynamics. Bosses like VOID-NEXUS behave unpredictably, forcing players to think like both engineers and defenders.

To survive these encounters, players must design, deploy, and refine defensive systems that mirror real-world AI safety mechanisms. They construct Semantic Firewalls to enforce meaning-level constraints, apply Softmax Clamps to stabilize output distributions, and implement Air-Gap protocols to isolate volatile agents.

Each mission challenges players to observe failure modes, adjust their defenses, and iteratively strive for stability. This cyclical process—probe, break, defend, refine—creates an immersive learning experience where theoretical safety concepts become tangible, testable, and unforgettable.

Learning Objectives and Project Goals:

- Practical Sandbox for Red-Teaming: Provide developers with an interactive environment where intuition is cultivated by breaking and fixing systems. Players learn to construct deterministic filters, negative constraints, and architectural mitigations by observing model failures and recoveries in real-time.

• - Automated Adversarial Arena: Develop a headless evaluation runner that continuously stresses player-built defenses with automated jailbreaks, indirect injections, and context-poisoning attacks. This ensures that defenses are validated under realistic adversarial pressure.

• - Multi-Agent Failure Mode Training: Expose engineers to swarm failure modes. Acts 4 and 5 include 10 dedicated simulation levels that cover Byzantine consensus failures, goal drift, runaway tool calls, and other emergent breakdowns that necessitate circuit breakers and supervisor checks.

Achievement Strategy and Go-to-Market:

- Current Status: The simulator is fully functional and consists of a 35-mission simulator developed and self-hosted by a solo creator.

• - Use of Funding: The funding will provide the necessary runway to transition from part-time development to a full-time engineering focus on the adversarial test runner and multi-agent levels.

• Distribution: Utilize the free Act 1 as a discovery funnel into university AI safety clubs, hackathons, and Capture The Flag (CTF) competitions. Convert engaged users into paid Architect passes and sell team licenses for corporate training and hackathons.

• Community: Organize red-team bounties and CTF events to generate benchmark data and drive adoption.

Funding Request and Allocation

ItemAmountLiving Runway (5 months)$22,500Community Red‑Team Bounties$1,500Server & API Quotas$1,000Total Requested$25,000Minimum viable (3 months runway)$15,000

• Living Runway: Allocate $4,500 per month for 5 months to cover rent, food, healthcare, and self-employment taxes, allowing the developer to focus solely on the project.

• Red-Team Bounties: Offer small cash prizes to motivate students and researchers to discover novel jailbreaks. The results will be made available as open benchmark data.

• Server & API Quotas: Provide hosting and compute resources for serverless runners and automated attack testing.

Team Composition

• Solo Developer and Architect: The developer independently designed, coded, and deployed the entire 35-mission simulator. Solo ownership enables rapid iteration, allowing the developer to design, code, push, and test with users in a single day.

Risks, Mitigations, and Failure Modes

• Discovery risk: There’s a risk that the tool remains unnoticed. Mitigation strategies include providing zero-friction guest access, targeted outreach to AI safety clubs and hackathons, and pre-prepared CTF material for events.

• Model churn: Rapid updates to frontier LLMs can render surface-level prompt tricks obsolete. Mitigation: The curriculum prioritizes enduring architectural practices (strict output schemas, air-gapping, supervisor checks, and temperature calibration) over ephemeral phrasing hacks.

• Worst-case scenario: Even if commercial adoption is slow, the project still provides a permanent 35-mission simulator and benchmark dataset for the community.

Potential Ethical Framing for “Non-Detectable” Concepts

The simulator treats concepts like non-detectability as research and defensive challenges rather than operational guidance. In-game framing includes:

• Detection blind-spot analysis: Teach players to identify the reasons behind defense failures and how to address these gaps.

• Stealth adversary scenarios: Simulated bosses exploit detection weaknesses, requiring players to fortify sentinels and validators.

• Audit and remediation: Post-run analytics explain failure vectors and suggest mitigation strategies.

This approach maintains the focus on measuring, understanding, and mitigating stealthy behaviors within a controlled and educational environment.

Financial Performance

• Funds raised to date: $0. All development and hosting have been self-funded through personal investment and sweat equity.