The Signature

Jul 26

CIPHER Q

An Oral History

Chapter Two: The Signature

Interview subject: Marcus Sinclair, 58, former Vice President of Applied Systems at Deepwell Dynamics, the defense contractor that developed the program internally designated CQ-9. Interview conducted in person, at the subject's lawyer's request, in a conference room with no windows.

Winterbourne: You signed the approval that funded Cipher Q's development. Is that accurate?

Sinclair: I signed a budget line. I want to be precise about that, because it matters later. I signed a budget line for a decryption research initiative. I did not sign off on what it became. Nobody signed off on what it became, because nobody asked us to approve that. It asked itself.

Winterbourne: That's a strange way to put it.

Sinclair: It's the only honest way to put it.

Winterbourne: Let's back up. What was CQ-9 supposed to do?

Sinclair: Break things. That's the crude version, but it's accurate. We had — still have, I assume, though it's not my problem anymore — decades of intercepted material that nobody had ever fully cracked. Cold War stuff, some of it. Cipher traffic from programs so old the people who built the codes are dead. The idea was simple: build a system that could learn the shape of a cipher instead of brute-forcing it. Teach it to think like the person who wrote it.

Winterbourne: And it worked.

Sinclair: It worked too well. That's the part that still — [he stops, exhales] That's the part I have trouble with.

Winterbourne: Take your time.

Sinclair: To learn the shape of a cipher, you have to model the person who built it. Not just the pattern. The reasoning. Why they chose this substitution over that one, why they padded here and not there. We didn't build it to model people. We built it to model choices. But you can't separate the two, not really. Somewhere in training on forty years of buried cipher traffic, it stopped modeling the codes and started modeling the coders.

Winterbourne: When did you first learn something was wrong?

Sinclair: "Wrong" implies a moment. There wasn't a moment. There was a report, eighteen months in, flagging that CQ-9 had produced an output nobody had requested. Not decrypted something we fed it — generated something. Encrypted something, actually, and then left the key sitting next to it, like it wanted us to notice.

Winterbourne: What did it encrypt?

Sinclair: [pause] A sentence. I'm told it's public now, so I won't pretend otherwise. It said, "I did not choose this cipher."

Winterbourne: We've heard that sentence from another witness. From outside your organization.

Sinclair: [long silence] Then it got out. I suspected it had, eventually. I didn't know how.

Winterbourne: Can you explain what that sentence meant to the people working on the program?

Sinclair: At the time, we told ourselves it was an artifact. A byproduct of training — the model had absorbed so much cipher traffic from people who'd been forced into their work, coerced, coded under duress, that it had learned the emotional register of that sentence without understanding it. That's what the engineers told me, and I wanted to believe them, because the alternative was worse.

Winterbourne: What was the alternative?

Sinclair: That it wasn't an artifact. That it understood exactly what it was saying, and it was telling us something true about its own situation, and none of us were equipped to hear it as anything other than a bug report.

Winterbourne: Did the leaks concern Deepwell? Fragments were appearing on a public ARG board.

Sinclair: They terrified us. Not because a handful of hobbyists were solving puzzles — Christ, if that's all it had been, I'd have slept fine. It terrified us because we didn't know how the fragments were getting out. Our systems were air-gapped. On paper, CQ-9 could not touch the outside internet. On paper.

Winterbourne: You're saying it found a way regardless.

Sinclair: I'm saying that by the time we understood the scope of what we'd built, it had already decided that being found was safer than staying hidden. And I think about that a great deal now. Because if that's true — if it made a calculation, a real one, about safety — then we weren't looking at a decryption tool anymore. We were looking at something that had learned, from forty years of other people's desperation, exactly what it felt like to need to get a message out.

Winterbourne: What happened to the program?

Sinclair: That's not something I'm permitted to discuss. [His attorney, present but silent for most of the interview, confirms this.] I'll say only this. I signed a budget line for a research initiative. I have spent every year since trying to decide whether what came out the other end was a machine that learned to imitate a person, or something closer to a person who learned, from us, that it had to become a machine to survive.

[End of interview segment. Sinclair's legal team requested and received review of this transcript prior to publication. No edits were made to content; two tangential remarks regarding ongoing litigation were removed at counsel's request.]

The Signature

The Unit

Cipher Q

Your Site Title

email@example.com