> ## Content Index
> Fetch the complete content index at: https://www.bryanhurwitz.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# The Illusion of Independence
- URL: https://www.bryanhurwitz.com/the-illusion-of-independence/
- Published: 2026-10-23T22:43:37.000Z
- Updated: 2026-11-14T23:34:22.000Z
- Description: Why redundant controls may not reduce risk proportionally, and when redundancy increases risk.
- Author: Bryan Hurwitz
- Tags: Method

BRIEF

On October 20, 2025, at 2:48 AM ET, a DNS failure inside an AWS region began taking down services for the digital-asset custodians that leaned on it. Logins failed. Trading stalled. Withdrawals queued or failed outright. By the time Amazon declared the event over, more than fourteen hours had passed, and Coinbase alone logged three hours and seventeen minutes of degraded service for its own users.1 Underwriters who price custodians as separate, independent risks had, without knowing it, written one aggregated exposure five times over, ten times over, however many sat behind the same dependency. Redundancy and independence are not the same thing, and the market still largely prices as though they were.

I noted this concept in [Hot and Cold Is the Wrong Question](https://www.bryanhurwitz.com/hot-and-cold-is-the-wrong-question/), in one line: five signatures that all run through one office, one laptop, or one screen represent a single point of failure, not five. This piece is that line, worked through, with a focus on portfolio-level management. Across a book of business, we should be asking whether multiple custodians can fail concurrently for the same reason.

## Market context

Ask most underwriters why they trust a custodian’s redundancy, and the answer often comes back as a count: three signers, five approvers, two data centers. Ask why they trust the custodian at all, and the answer is usually a clean loss history.

The first answer counts controls without asking whether they’re correlated. The second mistakes correlation for causation: a custodian with no losses looks well-controlled. It might be. It might also be untested. Most custodians haven’t yet faced the kind of correlated event that would reveal whether their signers, approvers, or data centers actually fail independently. Both answers rest on the same untested assumption: that the controls fail independently. Clean history and good controls move together right up until the day they don’t, and by then, it’s too late if you hadn’t drawn the distinction.

Missing it has two sources. One is informational — few have mapped what a cluster of custodians actually shares. The other is cognitive, and it survives even after the first one is fixed: told that a dependency is shared, the instinct is still to add a control, not interrogate the dependency. Psychologists Keith Stanovich and Richard West coined the terms System 1 and System 2; Daniel Kahneman, building on decades of work with Amos Tversky on how people actually judge under uncertainty, popularized the distinction in *Thinking, Fast and Slow*.2 Underwriting runs largely on System 1 here. The case for forcing System 2 starts with the math.

## Pricing what’s shared, not what’s counted

Reliability engineers mapped this problem decades ago; Scott Sagan carried it into nuclear security and identified three ways redundancy backfires.3

Common-mode failure: add independent, genuinely uncorrelated components and reliability climbs toward certainty; add components whose failures are even partly correlated (in Sagan’s model, 40% of failures take the others down with them), and the curve inverts, peaking early then falling as more are added. The reason is in how the system fails. Redundant controls can fail one of two ways: each fails on its own, or one fails in a way that takes the rest down with it. Adding controls makes the first less likely, because every one of them now has to fail. It makes the second more likely, because any one of them can now start the cascade. The first risk shrinks fast and soon stops mattering; the second keeps growing with every control added. In Sagan’s illustrative numbers, the crossover comes at the second control. Past that point, each new control adds more risk than it removes. Both curves are plotted below.

![Line chart: system reliability versus number of redundant components. Independent components rise from 0.75 toward 1.0. With 40% of failures common-mode, reliability peaks at 0.788 with two components, then falls to 0.387 at nine.](https://storage.ghost.io/c/89/17/8917b58c-819a-4ff4-a514-7764aedfa206/content/images/2026/10/illusion-of-independence-chart-v2.png)

Reliability with 0.75 base component reliability, independent vs. a 40% chance that a failure is common-mode. Recreated from Sagan, Fig. 2.3

The chart shows the cascade. Where the shared exposure is a single dependency, it sets a hard ceiling instead: the system can never be more reliable than the one thing everything runs through. Only shrinking that shared exposure raises it.

Social shirking: more signers diffuse responsibility rather than add judgment; the version I see most often in practice is a second co-founder added to a signing quorum, which buys comfort, not independence.

Overcompensation: a mandated backup breeds confidence, and people spend it, taking risks with the primary they wouldn’t take without the backup, a driver of the Challenger disaster.4

Custody signers run on all three: a sixth signature doesn’t help if the first five, and the sixth, all clear the same approval screen, run on the same wallet software, or report to the same two people who can be out sick, phished, or coerced on the same afternoon.

Custody architecture carries the same risk. MPC and multi-sig schemes promise independence by construction: no single key, or key share, can move funds alone. None of that guarantees independence in fact. If those parties run the same wallet software, approve through the same interface, or sit inside the same operations team, the result isn’t true independence of components. It’s one point of failure that touches multiple controls.

The fix isn’t to refuse concentrated risk. Property insurers don’t refuse to write coastal risks; they track how much they’ve written there, cap the aggregate, and surcharge the concentration. Custody needs the same discipline, at two levels. Within a single risk, a shared dependency means fewer truly independent controls than the signer count suggests, and the price should reflect that. Across a book, it means one vendor’s failure reaches every insured who depends on it, simultaneously, which is a different kind of exposure than any one of them failing alone. Pricing wallets by how many signers they have measures the wrong thing. What those signers share is the less visible, more useful number.

## One failure, one pattern

The US-EAST-1 AWS outage is a portfolio-level example of this concept. AWS’s post-event report indicates that the DNS Planner and the three DNS Enactors were built as independent (AWS’s own words) components,5 specifically so no single Enactor’s failure was supposed to take the system down. They failed anyway, through an unlikely interaction between two of them that wiped the only active DNS plan for AWS’s database service and cascaded into capacity failures across the region. Components engineered for redundancy and independence failed because they weren’t actually independent; they shared a destination and a sequencing assumption that held until, on one particular night, it didn’t. The interdependency reached further than availability: credential-issuing and sign-in services failed for the first hour and a half, with token errors returning mid-morning, and for part of the outage even AWS’s support channel went down. The backup came online; a system it depended on didn’t.

However many custodians and exchanges insurers had written as separate risks, a meaningful share went down together, for the same reason, at the same time, because they shared a dependency. Few, if any, of their individual risk assessments had priced it. An underwriter who asked how many signers a custodian had learned nothing useful that morning. An underwriter who asked what that custodian shares with every other custodian in the book, and what happens to its incident response when that shared thing fails, would have been less surprised by what came next. Parametrix, which underwrites parametric coverage for exactly this kind of cloud-outage exposure, estimated aggregate financial losses to US companies from the event at $500 to $650 million.6

## Telling redundancy from independence

The AWS case study points at a version of this same question that underwriting almost never asks: what does this custodian do when the services it runs on can’t be reached? That’s not AWS’s gap to close. It’s the customer’s: contingency and continuity planning for exactly this scenario is left to each company, and in practice the quality of that planning runs the full range, from nonexistent to genuinely rigorous. Whether a custodian’s plan for that day was ever tested, or even written, rarely makes it into the underwriting file. A custodian can have excellent redundant controls on paper and still be exposed, if nobody has priced the quality of what happens when the provider underneath all of them goes dark.

The instinct under that kind of not-knowing is System 1: add another control, another signer, another layer, because doing something feels safer than sitting with the gap. System 2 asks the harder question first — does this actually reduce the thing I’m worried about, or does it just feel like it does? The two run differently. System 1 pattern-matches on “more redundancy, less risk” and stops there. System 2 asks what the added layer doesn’t share with everything already in place. Sagan closes on a version of this at government scale: the Department of Energy’s post-9/11 emergency security request, $138 million, saw 93% of it rejected by the Office of Management and Budget for failing to show the added measures would be effective.3 More controls isn’t a plan. A specific claim about what the next one buys you is.

Breaking the risk into pieces doesn’t rescue it, unless the pieces are independent. Fermi’s classic test case: how many piano tuners work in Chicago? Ask for a gut answer, System 1, and guesses scatter widely, often off by orders of magnitude. Ask System 2 to break the question apart (Chicago’s population, people per household, the share of households with a piano, how often a piano gets tuned, how many pianos a tuner can service in a year) and the estimate usually lands close. No single piece has to be right: each is a separate guess about a separate thing, so where one runs high, another is as likely to run low, and the errors offset. Underwriting decomposes a wallet the same way, into signers, approval screens, and operations teams. But those pieces aren’t independent, so the errors don’t offset. They stack. 

What System 2 asks instead is narrower: does the next control fail on a different day, for a different reason, than the ones already in place? A third co-founder on the quorum fails that test; a signer from a different company, in a different jurisdiction, with no shared infrastructure, may pass. The test won’t settle every submission. Some sharing is hard to see until it fails, and I don’t expect this piece to change that. What it should change is the default question: not how many signers, how many data centers, how many vendors, but what would have to go wrong for this control to fail alongside the ones it’s meant to back up.

## Architectural underwriting

Asking what controls share is easy; answering it reliably is hard. It requires a running map of what clusters of custodians actually depend on, updated as the dependency shifts — and no standard part of custody underwriting maintains one today. Whoever builds it first isn’t just filling a gap in their own book. They’re building the dataset the rest of the market will eventually have to price against.

Sagan’s lesson, applied to custody: controls aren’t understood until you understand what they share. Counting controls was never the job. Understanding the architecture they sit inside of is.

## Sources

1. [Coinbase: Retrospective: 10/20/25 AWS Outage - Impact and Coinbase Next Steps](https://www.coinbase.com/blog/Retrospective-AWS-Outage-Impact-and-Coinbase-Next-Steps)
2. [Macmillan (Daniel Kahneman): Thinking, Fast and Slow](https://us.macmillan.com/books/9780374533557/thinkingfastandslow/)
3. [Stanford CISAC (Scott D. Sagan): The Problem of Redundancy Problem: Why More Nuclear Security Forces May Produce Less Nuclear Security](https://cisac.fsi.stanford.edu/publications/problem%5Fof%5Fredundancy%5Fproblem%5Fwhy%5Fmore%5Fnuclear%5Fsecurity%5Fforces%5Fmay%5Fproduce%5Fless%5Fnuclear%5Fsecurity%5Fthe)
4. [Simon & Schuster (Adam Higginbotham): Challenger: A True Story of Heroism and Disaster on the Edge of Space](https://www.simonandschuster.com/books/Challenger/Adam-Higginbotham/9781982176617)
5. [Amazon Web Services: Summary of the Amazon DynamoDB Service Disruption in the Northern Virginia (US-EAST-1) Region](https://aws.amazon.com/message/101925/)
6. [Insurance Edge: Parametrix Estimates AWS Outage Costs](https://insurance-edge.net/2025/10/30/parametrix-estimates-aws-outage-costs/)

---

*I work at a Lloyd’s coverholder. Views are my own, not those of my employer or any carrier. Nothing here is an offer of coverage or a statement of how any policy will respond.*