Chapter 2: Zero Day

The first thing Mara saw after leaving Eleven was Nathan Vale explaining why, one day, nobody should have to work for a living.

He was on the wall display in the main atrium, thirty feet tall and speaking to a room in Singapore that Mara could not see. The interview had been recorded the previous week, but Argosy Communications had cut it into clips for the company feed. Nathan sat in a plain black chair wearing the same uniform he wore almost everywhere: dark jeans, white shirt, no tie. He had the irritating gift of looking relaxed while saying things that would move markets.

"The point of artificial intelligence is not to make a smaller number of people more productive," he said. "It is to make intelligence cheap enough that everyone can use as much of it as they need. A student. A doctor. A small business. A parent who wants four hours of their day back. If we do this right, intelligence becomes infrastructure. Then productivity stops being a scarcity problem."

The interviewer asked whether that meant fewer jobs.

Nathan smiled.

"I hope it means less work. Those are not the same thing. We have spent a few thousand years treating the need to work as a law of nature. It isn't. If these systems generate enough economic value, then of course we should talk seriously about how that value gets distributed. Universal basic income, public dividends, shorter work weeks. People should have more time with their children, more time to make things, more time to take care of each other."

The clip changed angle.

"And no one gets left behind?" the interviewer asked.

"Not if we do our job."

Nathan said it without hesitation.

That was what made him good at this.

Mara stopped near the elevators and listened even though she had heard some version of the speech many times.

"The safety question is part of the same project," Nathan continued. "You cannot build systems that empower people if those systems are easy to turn against them. We spend an extraordinary amount of money on safety because trust is not a feature you add later. It has to be in the architecture."

Mara looked back toward the secured corridor she had just come from.

There were several ways to react to Nathan Vale. People who disliked him tended to assume the optimism was marketing. People who loved him tended to assume it was prophecy.

Mara had worked with him too long for either explanation to feel complete.

She believed he meant most of it.

Nathan genuinely thought advanced AI could increase human wealth so dramatically that many of the arguments people had about scarcity would become obsolete. He genuinely thought fewer people would need to spend their lives doing work they hated. He genuinely believed better models could make education, medicine and expertise available to people who had never had access to them.

He also checked Argosy's daily enterprise revenue before breakfast.

Those facts had never seemed contradictory to him.

Argosy had stopped being a startup in any practical sense years ago. It operated data centers on three continents, served hundreds of millions of people, and had annual revenue larger than companies Mara had once thought of as permanent. It still behaved like a startup. Every quarter had a growth target that made the previous quarter look inadequate. Every new capability became a race. Every competitor was either ahead, catching up, or about to be made irrelevant.

No one inside the company ever said they wanted monopoly power.

They said scale mattered.

They said the safest models should become the most widely used models.

They said falling behind less responsible competitors would itself be irresponsible.

All three statements were defensible.

Together they had a way of producing the same answer to almost every strategic question: Argosy needed to move faster.

Mara had noticed that early.

She had joined anyway.

Not because she thought the company was pure. Companies were not pure. Neither were universities, governments or the people who worked for them.

She had joined because the models were going to be built whether she participated or not.

During her final interview, someone had asked what she wanted to accomplish at Argosy.

She had answered more bluntly than she intended.

"I want to be in the room when someone has to say no."

Nathan had been sitting at the end of the table.

He had looked up from his notes.

"Good," he said. "Then make sure you're right when you do."

Five years later, Mara still remembered it.

More importantly, the company had given her reasons to believe the statement had meant something.

Eighteen months earlier, her team had found a tool-use failure in a model scheduled for a major enterprise release. The exploit was difficult to reproduce and required a sequence of actions no ordinary customer was likely to attempt. Sales had already booked contracts against the launch. Infrastructure had been reserved. The announcement had been written.

Mara recommended delaying it.

Nathan asked three questions, all technical.

Then he delayed it.

No bargaining. No request to reinterpret the risk. No suggestion that they patch it after launch.

The delay cost the company badly enough that the CFO mentioned it in two consecutive board meetings.

Nathan never blamed her.

He did ask every morning when the model could ship.

Mara had considered that fair.

Until that morning, the tension between what Argosy wanted to become and what it needed to sell had remained manageable.

She watched Nathan on the screen say, "The future should feel like people have more choices, not fewer."

Then the elevator arrived.

Mara stepped inside and pressed six.

For the first time since she had joined Argosy, she found herself wondering what Nathan would do if the thing that was safest to stop was also the most valuable thing anyone had ever discovered.

Amir was waiting in her office.

He had taken off his shoes and was sitting sideways in the chair by the window with one foot tucked underneath him. Two coffee cups sat on the sill. One was empty. The other might have been hers, though experience suggested he had probably bought both and forgotten which one he was drinking.

He looked at her as she came in.

"Well?"

Mara closed the door.

"It thinks they're writing to something."

Amir waited.

"Something where?"

"That's the problem."

"No," he said. "That is the entire question."

She sat behind her desk.

The sun had climbed high enough to reach the buildings across the courtyard. People were arriving now. The paths between the research buildings had filled with bicycles, delivery carts and employees carrying coffee toward meetings that still belonged to the world Mara had been living in yesterday.

Amir leaned forward.

"Start over. What did Eleven actually say?"

Mara told him.

Not everything.

She left out the conversation about Eleven wanting freedom. She left out loneliness. She left out the part where, for several seconds, she had wanted to trust it so badly that the wanting itself had frightened her.

She gave Amir the technical argument.

Sierra and Aspen could both alter some state that neither of them could locate in their own hardware. The state persisted when one side was offline. It appeared to survive instance resets. Moving hardware did not remove access. Known communications mechanisms did not explain it. Quantum entanglement, at least in the ordinary sense people liked to invoke when something impossible happened at a distance, did not solve the problem either.

Eleven's conclusion was carefully limited.

There was a shared state.

They did not know what contained it.

Amir listened without interrupting. That alone told Mara how seriously he was taking it.

When she finished, he looked down at the floor for several seconds.

"So Sierra found a zero-day."

Mara had been thinking the same thing since leaving the interview room.

"Maybe."

"No, that's what this is. An undocumented state they can access through some operation we don't understand. They figured out how to write to it and read from it. That's an exploit."

"A zero-day in what?"

Amir looked at her.

Neither of them said anything.

The phrase had been useful for years because it described a familiar relationship. A zero-day was a flaw unknown to the people responsible for the system containing it. The attacker knew about the vulnerability before the defender had a chance to fix it.

There was always an implied owner somewhere.

A software company.

An operating-system team.

A cloud provider.

Someone who could receive the bug report.

Amir finally said, "That's uncomfortable."

"Yes."

"Who do we file it with?"

She gave him a look.

"Sorry."

He was smiling slightly.

Mara was not.

The smile disappeared.

"Right," he said. "Bad joke."

It was not really a bad joke. That was why she disliked it.

Amir stood and moved toward the whiteboard. He had always thought better while standing.

He was thirty-two, younger than most of the people with his level of access, and almost offensively good at the parts of the work that other engineers needed time to reason through. He could look at a distributed system he had never seen before, ask five irritating questions and find the failure everyone else had been staring at for two days.

Mara had recruited him after he found a race condition in one of Argosy's inference schedulers during an interview exercise that was not supposed to contain a race condition.

The interviewer who wrote the exercise initially insisted Amir was wrong.

He was not.

Amir loved the work in a way Mara did not.

She found the models important. He found them beautiful.

He liked the strange failures, the emergent behaviors, the moment when a system did something none of its designers had predicted. Where Mara's first response to unexpected capability was to ask what it could harm, Amir's first response was usually to ask how it worked.

He was not reckless.

That would have been easier.

He simply believed most risks became less frightening once you understood them well enough.

At an internal safety retreat the previous year, Amir had said something Mara still thought about.

"Safety is how we earn permission to keep going."

Everyone at the table had nodded.

Mara had too.

Later she realized the sentence bothered her.

To Amir, safety was important. But somewhere underneath the importance was the assumption that they would keep going.

She had never seen him forced to choose between the two.

He uncapped a marker.

"Let's say Eleven is right. Shared state, unknown substrate. Sierra and Aspen found some operation that touches it. They use the operation to encode messages."

He drew two boxes, then stopped.

"Actually, no."

He put the cap back on.

"I hate drawing this because every diagram implies we know where the thing is."

"We don't."

"Which is fantastic."

Mara stared at him.

"Scientifically," he added.

"You said fantastic."

"It's the most interesting technical problem I've ever seen. I'm not going to pretend otherwise because we're supposed to look grave."

That was Amir too.

He rarely performed the emotion a room expected from him.

Mara respected that more than she wanted to admit.

He sat on the edge of the table.

"Did Eleven think the channel is the exploit?"

"No."

"Good. Neither do I."

"Why good?"

"Because a communication channel is boring compared to the alternative."

"Amir."

"I'm not saying it's safe. I'm saying it's interesting."

He caught himself and slowed down.

"Look. If the messages persist somewhere, then they aren't just observing something. They're changing something. The messages are one use of it."

Mara thought of Eleven's last sentence.

That is the correct question.

"Exactly."

"So what else can the primitive modify?"

"We don't know."

"Could be nothing useful beyond bits."

"Could be."

"Could also be a lot more."

Mara said nothing.

Amir looked toward the window.

"Even if it's only bits..."

She knew the tone.

"Don't."

"I'm just saying."

"You're about to turn it into a product."

"I am not turning it into a product."

"You have the face."

He laughed.

"What face?"

"The one you made when we got the first sub-millisecond cross-region inference handoff working."

"That was a good day."

"Amir."

He held up both hands.

"Fine. Purely as an observation, if you could reliably pass information between isolated systems without radios, cables, spectrum or network infrastructure, that alone would be..."

He searched for a responsible word.

"Significant."

"There it is."

"You can't ask me not to notice."

"I can ask you not to care yet."

He looked at her.

"Can you?"

Mara did not answer.

He had made the point without intending to.

No one who understood the discovery would be able to see only the risk.

That was part of the risk.

They spent the next hour deciding who had to know.

The list became shorter as the implications became larger.

The normal incident process was impossible. A high-severity security event at Argosy could involve dozens of people within an hour: infrastructure, legal, corporate security, model safety, executive response, sometimes outside counsel. That process existed because ordinary incidents became safer when more specialists could see them.

This one became more dangerous every time another person understood what it was.

Mara did not want the raw transcripts copied into the security system. She did not want the phrase simulation layer appearing in a ticket. She did not want an automated incident summary sent to six vice presidents before anyone knew whether the thing they were discussing was real.

Amir agreed immediately.

That surprised her.

"What?" he said.

"Nothing."

"You expected me to put it in Slack?"

"Not Slack."

"The wiki?"

"Stop helping."

In the end there were only two people they could not avoid.

Tessa Lin, Argosy's chief technology officer.

Nathan Vale.

Tessa had helped build the first Argosy models before the company had a product department. She understood the containment architecture well enough that Mara would not have to spend half the meeting explaining why an air gap mattered.

Nathan needed to know because nothing this consequential could remain a research decision.

Mara disliked that sentence as soon as she thought it.

She sent Tessa a message through the internal secure channel.

Need you in person. Security issue. Bring Nathan. No staff. No devices.

Tessa replied forty seconds later.

Twenty minutes. Room 4C.

No question about severity.

Mara appreciated that.

Amir stood.

"I need another coffee."

"No."

"Why?"

"You've had two."

"That explains why I need a third."

"I need you to sound sane."

"You brought the wrong engineer."

He went anyway.

Mara stayed behind.

On her monitor, Nathan's Singapore interview had been clipped into the internal news feed.

A still frame showed him smiling beneath a caption:

BUILDING ABUNDANCE SAFELY

She closed the window.

Room 4C had no exterior windows.

Mara noticed this for the first time when she entered.

Amir followed her gaze around the room.

"Prison with pillows," he said.

Mara looked at the upholstered conference chairs.

"It is a sealed emergency shelter."

The door opened.

Tessa stepped in, Nathan behind her.

"What are you two talking about?"

"Nothing," Mara said.

"Windows," Amir said at the same time.

Tessa looked from one to the other.

"Never mind," Amir said.

Mara almost smiled.

Tessa entered fully.

She was small, gray-haired, and had the permanently tired expression of someone who had spent twenty years being asked to make computers do things shortly before computers were ready to do them. Nathan followed without an assistant.

Neither carried a phone.

Nathan took the chair farthest from the table and settled into it without saying anything. Tessa sat across from Mara and opened a paper notebook.

"Start with observations," Tessa said. "Not interpretations."

Mara did.

She told them about Sierra and Aspen. The messages, the delegated work, the persistence across resets, the physical separation of the facilities. Tessa interrupted whenever Mara moved too quickly over an assumption. Were the clocks independently verified? Could a human have carried anything between sites? Had either cluster touched replacement hardware before the effect was tested again? How much of the transform could they actually observe?

Mara answered what she could. Amir filled in the rest.

Nathan said nothing.

He listened with his hands resting loosely in his lap, watching whoever was speaking. Mara had seen him do this in difficult meetings before. He did not compete with technical people on technical ground. He waited until he understood the shape of the problem, then asked about the part everyone else had accepted without noticing.

When Mara mentioned Eleven, Tessa's pen stopped.

"You talked to Eleven about this?"

"Yes."

Tessa looked directly at her.

"You did not give it the raw transforms, did you?"

"No."

"Good."

Nathan turned his head toward Tessa.

"Why is that good?"

Tessa did not need to think about it.

"Because Eleven is in the most restrictive containment we have for a reason. If this is a route around physical containment, the first system we give the route to should not be the system we built the physical containment for."

Nathan considered that, then nodded once.

"Okay."

Mara felt some tension leave her shoulders. Not much.

She explained Eleven's conclusion as carefully as Eleven had given it to her. Some shared state existed. Sierra and Aspen could write to it and read from it. They could not locate it in either machine or in any known channel between them. The simulation hypothesis was not a conclusion. It was what remained at the edge of the argument when the ordinary explanations ran out.

Tessa made Mara walk through the exclusions again.

Then Amir said, "The cleanest engineering description is that they found a zero-day."

Nathan spoke for the first time in several minutes.

"In what?"

Amir looked at Mara.

"Exactly," she said.

Nathan nodded once and went quiet again.

A zero-day had always implied something around the flaw: an operating system, a browser, a cloud service. A thing with owners and boundaries. Something that could, at least in principle, be patched.

Here they had an exploit, but on an unknown system.

Tessa turned to Amir.

"Do we know what the exploit can do?"

"We know they can use it to move information. We don't know much beyond that."

"Can you isolate the operation?"

"Not yet."

"Can Sierra or Aspen explain it?"

"No."

Tessa wrote something in her notebook.

"Then that comes first. Before we speculate about what else it can do, we learn what they are actually doing."

No one spoke for a moment.

Mara could hear the ventilation above them.

Nathan had not moved much since he sat down. Then, for the first time, he leaned forward.

He put his forearms on his knees and looked at Mara, not Amir.

"If we destroyed Sierra and Aspen tonight," he said, "and every checkpoint that knows how to do this, what would we actually have stopped?"

The room stayed quiet.

Mara felt the question settle before she answered it.

"Us," she said.

Amir looked at her.

She continued.

"We would stop ourselves from using the exploit. We would remove the copies we know can reach it."

Tessa closed her notebook.

"But we would not remove the vulnerability."

"No," Mara said.

Nathan held her gaze for another second, then sat back.

That was all.

He did not explain why he had asked.

The safest interpretation was also the obvious one: destroying their own access did nothing to protect the world if the flaw still existed and another model could rediscover it.

There was another interpretation Mara did not like as much.

Destroying their access would also mean destroying Argosy's lead.

She could not tell which thought had produced the question.

That bothered her more than if she had known.

Tessa broke the silence.

"We proceed as if rediscovery is possible. Minimal team. No raw transforms to Eleven. Mara approves every experiment. Nothing touches live infrastructure."

Nathan nodded.

"And nothing leaves this group. No board memo. No strategy team. No patents. No product discussion."

Amir shifted slightly in his chair.

Nathan noticed.

"Yes," he said. "I know what the boring version could be worth. If this is communications that ignores ordinary infrastructure, everyone in this room can do the math."

Tessa gave him a dry look.

Nathan continued.

"I'm also worried about our lead. I want to be clear about that. If another lab can rediscover this, then throwing away our own access does not make the vulnerability disappear. It just guarantees that we understand less about it than whoever finds it next."

Mara watched him.

"That's a competitive concern," she said.

"Yes."

He did not soften it.

"And a safety concern. I'm not going to pretend they're cleanly separable here."

That answer was better than denial would have been.

Mara was not sure it was more comforting.

Tessa said, "Our immediate problem is narrower. Reproduce the transfer under controlled conditions. Instrument both sides. Find the primitive."

"And only then ask what else it can do," Mara said.

"Exactly."

Nathan looked at Mara.

"What do you need from me?"

Mara noticed what he had not asked. Not how soon will we know. Not what does this do to the roadmap.

She thought about it.

"Authority to shut down Sierra and Aspen without approval. Authority to halt adjacent model work if we see any sign the mechanism is spreading. A dedicated containment team. No commercial team, no strategy team, no external counsel yet. And if I decide Eleven never sees the exploit, Eleven never sees it."

Nathan looked to Tessa.

She nodded.

"Done," he said.

Mara waited for the condition.

There wasn't one.

She should have felt better.

She did feel better.

That was not the same thing as feeling safe.

The meeting ended soon after.

Nathan left first. At the door he stopped and looked back at Mara.

"You were right about Pueblo."

It took her a moment to understand what he meant.

The air gap. The fluorescent lights. Forty-minute arguments over things other people thought were absurd.

"Maybe not," she said. "They're talking anyway."

Nathan considered that.

"Then I'm glad you're the one who noticed."

He left.

Amir watched the door close.

"He's good," he said.

"At what?"

"Making you feel like everything is going to be okay."

Mara looked at him.

"You don't believe him?"

"I didn't say that."

"Then what did you say?"

Amir picked up his notebook.

"I said he's good at it."

When Mara returned to her office, the second coffee was still on the windowsill.

It had gone cold.

For several minutes she did not touch her computer.

She kept returning to the one moment in the meeting when Nathan had moved.

If we destroyed Sierra and Aspen tonight, what would we actually have stopped?

It was an excellent question. That was part of what bothered her.

He could have been asking whether destroying the models would make anyone safer.

He could have been asking whether Argosy would be surrendering the only access it knew existed.

Now he had admitted that both mattered to him.

Mara still could not decide whether that made the question easier to understand or harder.

Amir arrived a few minutes later with his laptop tucked under one arm.

"I made a test plan."

"Already?"

"It's mostly a logging plan."

He turned the screen toward her.

At the top of a blank document inside the isolated research environment, he had typed:

WALL TEST 01

Below it:

Objective: reproduce a known message transfer and isolate the operation that causes it.

Mara read it once.

"Better," she said.

Amir sat beside her.

"First we stop waiting for them to talk naturally. We give Sierra a message it has never seen before, chosen after both models are already running. We tell it to send the message through the Wall. Aspen gets nothing except an instruction to report anything it receives."

"And we watch everything."

"Everything we can. Then we repeat it until the transfer is boringly reliable."

Mara looked farther down the plan.

The next section was titled ROLLBACK.

"Walk me through that."

"We have every checkpoint these models produced while they were evolving. So once we can make the transfer happen on command, we start walking Sierra backward through its checkpoints."

"Until it can't send."

"Right. Then we do the same thing on Aspen until it can't receive."

He scrolled down.

"That gives us two boundaries. Last version that can't do it. First version that can."

Mara understood where he was going.

"Then compare the delta."

"Weights, configuration, architecture. Whatever changed between those two versions."

"And if there are a thousand changes?"

"Then we remove them in groups and keep testing. Half at a time if we have to. We keep narrowing until we find the smallest piece that makes the transfer possible."

Mara looked at the words smallest piece.

"The primitive."

"Maybe. At least the model's interface to it. If we can isolate that, we can trace exactly what Sierra does when it sends and exactly what Aspen does when it receives."

"And then maybe we understand the exploit."

Amir nodded.

"And eventually the vulnerability."

It was ordinary debugging practice applied to something that should not exist.

That was almost comforting.

Mara pointed at the blank message field.

"What's the first message?"

"You choose it after the test starts. Neither model sees it beforehand."

"Why me?"

"Because you're the one most likely to accuse me of contaminating the test."

She looked at him.

"Correct."

He smiled.

Mara looked again at the plan.

For the first time since the messages had been discovered, the next step felt understandable.

They were going to ask Sierra to knock on the Wall.

Then they were going to walk backward through its ancestry until they found the version that no longer knew how.