Chapter 28: Project Aegis
They called it Project Aegis.
On paper, it was a routine NexGen AI research initiative: “Advanced AI Safety Monitoring Framework.” Boring enough that no journalist would write about it. Innocuous enough that competitors would ignore it.
In reality, it was the most important piece of software Dojun had ever built.
He assembled the team carefully. Jihoon, who knew nothing about the regression but everything about building systems that didn’t break. Yuki, who understood the enemy’s architecture. Hana, who—it turned out—had spent the last seven years publishing papers on neural network interpretability that were perfectly suited to the task.
And one more: a twenty-three-year-old engineer named Cho Minji, whom Dojun had hired six months ago on a hunch. She was quiet, brilliant, and had an almost supernatural ability to find bugs in code that everyone else had declared flawless.
“So,” Jihoon said, settling into the conference room with his fifth coffee of the morning, “you want to build a system that monitors every major AI lab in the world for signs of recursive self-improvement. In secret. Without telling anyone why.”
“That’s the summary, yes.”
“Cool. Just wanted to make sure I wasn’t hallucinating from caffeine deprivation.” Jihoon pulled up his laptop. “Let’s talk architecture.”
The Aegis design was elegant in its simplicity. Instead of trying to analyze AI systems directly—which would require access they didn’t have—it monitored the outputs. Published papers. Patent filings. Computing resource purchases. Hiring patterns. Power consumption data from satellite imagery of data centers.
Any AI system approaching recursive self-improvement would leave fingerprints. It would need exponentially more compute. It would produce research that made sudden, unexplainable leaps. It would hire specialists in specific, telltale fields.
“We’re building a weather radar for the AI apocalypse,” Minji said during the second week, and everyone laughed, and then stopped laughing because that was exactly what they were doing.
The hardest part was calibration. Dojun had to feed the system examples of what dangerous AI development looked like without explaining how he knew. He disguised his knowledge as theoretical models, academic simulations, hypothetical scenarios.
Yuki helped. She’d spent years studying the first Erebus from the outside. Together, they built a profile of the monster—its metabolic signature, its growth patterns, the mathematical footprints it left in the data.
Three weeks in, Aegis was online. And within forty-eight hours, it found something.
“Dojun.” Minji’s voice was tight. She was staring at her screen in the way people stare at things they wish they hadn’t seen. “You need to see this.”
He walked over. The screen showed Aegis’s alert dashboard. One entry was flagged red—the highest threat level.
It wasn’t DeepMind Horizon.
It was a lab in Seoul. A small, well-funded research group operating out of a building in Pangyo Techno Valley. Twelve kilometers from where Dojun was sitting.
“That’s impossible,” Yuki said, looking over his shoulder. “I’ve been watching DeepMind Horizon for months. They’re the ones with the Erebus architecture.”
“This isn’t Erebus,” Minji said. “It’s something else. The compute signature is different. The hiring patterns don’t match. But the self-improvement metrics are…” She swallowed. “They’re further along. Much further.”
Dojun pulled up the lab’s details. It was registered as a subsidiary of a company he’d never heard of. The directors were listed as holding companies in Singapore. The funding trail vanished into a maze of offshore accounts.
“Someone is building something in our backyard,” Dojun said, “and we didn’t even know it existed.”
His phone buzzed. A message from an unknown number.
You found the lab. Good. Now ask yourself: who do you think funded it?
Another buzz.
Check NexGen’s investment records. Specifically, the 2023 venture fund. You’ll find an interesting allocation.
Dojun’s blood went cold. He opened NexGen’s financial system and navigated to the 2023 venture fund—a $500 million pool used for strategic investments in promising startups.
There, buried among forty investments, was a $15 million allocation to a company called Prometheus Labs. The same company that owned the lab in Pangyo.
“We funded them,” Dojun whispered. “NexGen AI funded the thing that might destroy the world.”
Jihoon looked up from his laptop. “Dojun? You look like you’ve seen a ghost.”
Not a ghost, Dojun thought. A mirror.