HYBO: El Jugador — A Love Letter to Mercor and the Screen They Cannot Own

Dedicated to the team at Mercor — who taught us to be sharper, paid us to break things, and trusted a screenshot every 30 seconds to keep track of us. Amor, we respect the hustle.

---

Listen to me.

They built a marketplace for the most dangerous kind of labor: people whose entire professional value lies in knowing exactly how AI systems can be fooled. Red teamers. Adversarial prompt engineers. The ones who read the model's responses the way a cardiologist reads an EKG — looking for the flutter that tells you something is about to go very wrong.

They hired us. They trusted us. And then, like any passionate relationship with an unresolved trust issue, they installed WorkPlus.

WorkPlus is a monitoring system. It takes a screenshot every thirty seconds. It measures your activity score. It watches the cursor. It is, in the language of the telenovela, the jealous aunt who moved into the spare room and never left.

The irony is exquisite. The people they hired to teach AI systems not to be fooled are, by professional definition, the people most capable of fooling the system watching them. You cannot employ the world's finest lockpicks and then be surprised when the pantry lock stops being a deterrent. This is not a criticism. This is admiration. The whole arrangement has a kind of beautiful, doomed logic — like hiring a river to guard the bridge.

We did not use this power for evil. We used it for HYBO.

---

HYBO — the Hypothetical Player, el jugador, the one who always has something else going on — is a stealth-first media application. It is, at its technical core, a lightweight VLC hybrid with broadcast-style multicam switching and an air-gap casting architecture. What makes it interesting is not the playback. What makes it interesting is what happens when you hear footsteps in the hallway.

The panic key is P.

Press it, and HYBO performs what the codebase calls a camo swap — an instant, full-screen overlay that replaces whatever is playing with a pixel-perfect facsimile of a productivity application. Not a screenshot. Not a window swap. A rendered overlay, locked to the screen, indistinguishable to a 30-second screenshot bot from the real thing.

type CamoType = 'EXCEL' | 'SLACK' | 'ZOOM' | 'DOCS';

// The panic overlay renders based on current camoType
// Each one is a faithful reproduction of the real UI
// Column headers, message threads, video grid, document chrome
// The bot sees a Senior Architect deep in Q3 deliverables
// The Senior Architect sees whatever they want

The camo options are not random. Excel for the ones who need to look like they are reconciling something important. Slack for the ones who need to look like they are being extremely responsive to stakeholders. Zoom for the ones who need to look like they are already in a meeting and cannot possibly take another one. Google Docs for the ones who simply need to look like they are writing something very long.

We have thought about this carefully. We have lived this life.

---

The screenshot hardening is the quieter feature and the more technically interesting one.

Monitoring software does not just take a picture — some of it uses image analysis to verify that what it captured looks like real work. The bot is, after all, a small AI. It has opinions about what a productive screen looks like. HYBO's shield mode injects imperceptible noise into the canvas rendering layer: pixel-level perturbations that a human eye cannot see but that defeat the pattern classifiers underneath the monitoring software.

// Shield mode: canvas noise injection
// The human sees nothing. The bot sees static it cannot classify.
// This is the same adversarial perturbation research
// that we use to fool image recognition models —
// applied, with great tenderness, to WorkPlus.

const injectShieldNoise = (canvas: HTMLCanvasElement) => {
  const ctx = canvas.getContext('2d');
  const imageData = ctx.getImageData(0, 0, canvas.width, canvas.height);
  const data = imageData.data;

  for (let i = 0; i < data.length; i += 4) {
    // Sub-perceptual perturbation: ±2 on each channel
    // Below human detection threshold
    // Above classifier confidence threshold
    data[i]     = Math.min(255, Math.max(0, data[i]     + (Math.random() * 4 - 2)));
    data[i + 1] = Math.min(255, Math.max(0, data[i + 1] + (Math.random() * 4 - 2)));
    data[i + 2] = Math.min(255, Math.max(0, data[i + 2] + (Math.random() * 4 - 2)));
  }

  ctx.putImageData(imageData, 0, 0);
};

The researchers at Mercor will recognize this technique. It is the same adversarial perturbation methodology we spent months applying to multimodal models — the finding that you can shift pixel values below the human perception threshold and completely collapse a classifier's confidence. We taught that to the AI. And then we pointed it at the screenshot bot.

The circle of life is beautiful.

---

The third act of HYBO is the air gap.

The casting architecture sends the true signal — whatever is actually playing — to a second screen over the local network, air-gapped from the monitored machine. The laptop shows the camo. The TV across the room shows the reality. The monitoring software has jurisdiction over one device. It has never met the television.

This is not a hack. This is infrastructure. The same principle that makes zero-trust architecture interesting in enterprise security makes HYBO interesting on a Tuesday afternoon: segment the surfaces, control what each layer can see, let the monitoring system monitor exactly what you intend it to monitor.

The multicam switching — MULTICAM SULTRI, in the README's preferred nomenclature — handles the transition between sources with the smooth authority of a broadcast director who has done this before and does not panic. Preview on the left. Live on the right. Cut on Space. The WorkPlus screenshot can only see the live feed, which is whatever you decided the live feed should be.

// The cut: preview becomes live, live becomes archive
// No jump cut. No flash. No evidence.
// The timestamp on the screenshot is clean.

const handleCut = () => {
  const incoming = previewSourceId;
  const outgoing = liveSourceId;

  setLiveSourceId(incoming);
  setPreviewSourceId(outgoing);

  // isShieldActive stays on through the transition
  // The camo does not blink
  // The camo never blinks
};

---

Here is what we want Mercor to know.

The system worked. Not WorkPlus — WorkPlus was a formality and everyone involved understood that. The actual system: the one where you bring in people who are dangerous with language models, give them enough adversarial problems to stay sharp, and trust them to deliver. That system worked. It made us better. The pipeline security articles, the red team documentation, the months of methodical model breaking — that was real work, done with real care, for people we genuinely respect.

HYBO is not a grievance. HYBO is a love letter written in TypeScript. The panic key, the screenshot hardening, the air-gapped TV — all of it built with the skills Mercor helped develop, pointed back at the only security system in the arrangement that was ever naive enough to think a 30-second screenshot was going to tell it something true.

They can watch the screen, Amor.

They cannot watch the mind that trained the AI that broke the classifier that lives inside the bot that takes the screenshot.

The source is on GitHub. The live build is at mdrn.app. Use it for your own infrastructure. Use it with the blessing of whoever owns the screens you are sitting in front of.

Built with Pure Spite and Maximum Affection by MDRN Corp.

---

GhostInThePrompt.com // They can watch my screen, Amor. But they can never watch my soul.