Simon Willison recently observed that cybersecurity now looks like proof of work, and the framing clarifies something that has been building for years without a name that fit.
In Bitcoin, proof of work means every block added to the chain required a demonstrable expenditure of computational resources. The proof is expensive to produce and cheap to verify. What makes it effective as a consensus mechanism is that attacks require outspending the honest participants, which at scale becomes economically irrational; the expenditure itself is the security guarantee.
The security analogy runs close. Patch your systems this month and you are not safe for the year; you need to patch them again next month, and every month after that. Monitor your logs today and that coverage expires by tomorrow’s alerts. Train your staff against phishing this quarter and the training expires before the threat landscape does. The work of staying secure is continuous, mandatory, and never concludes. There is no endpoint, no proof of completion that carries forward.
Proof of Work as an Actual Security Mechanism
Before examining why this is a structural problem, it is worth noting that proof of work has a genuine 30-year history as a targeted security tool, and in that context it works well.
Adam Back’s Hashcash, proposed on the cypherpunks mailing list in 1997, is the earliest practical instance. The mechanism required email senders to include a header containing a partial SHA hash collision: a string whose hash starts with a configurable number of zero bits. Finding such a string requires real computation. For a legitimate sender dispatching a hundred emails per day, the per-message cost is negligible. For a spammer sending a hundred million messages, the aggregate CPU cost becomes prohibitive.
The mechanism did something subtle worth isolating: it required no centralized authority. The sender computed the proof, the recipient verified it, and the asymmetry was cryptographic rather than institutional. Satoshi Nakamoto cited Hashcash directly in the Bitcoin whitepaper when describing the mining mechanism, which is one of the more interesting intellectual migration paths in computing history. The Dwork-Naor paper from Crypto 1992, “Pricing via Processing or Combatting Junk Mail,” independently explored the same idea earlier but never reached practical deployment at scale.
The principle transferred well to other narrow security problems. Password hashing algorithms like bcrypt, scrypt, and Argon2 are deliberately expensive to compute, specifically to make brute-force cracking prohibitive. A legitimate user pays a cost measured in milliseconds; an attacker checking billions of passwords pays a cost measured in months of GPU time or thousands of dollars in cloud spend. Cloudflare’s Turnstile browser challenge system includes computational puzzles that browsers solve invisibly in the background, making bot operators pay real CPU costs per request while legitimate users experience nothing.
In each of these cases, PoW works for the same structural reason: the attack is high-volume and automated, the cost per legitimate use is low, and the total attacker cost scales multiplicatively with attack volume. The cost math favors the defender.
Where the Analogy Becomes a Problem
Organizational security defense does not share those structural properties. The work requirement grows without an upper bound, and there is no feedback mechanism analogous to Bitcoin’s difficulty adjustment that caps it at a sustainable level.
Consider what a reasonably security-conscious engineering team needs to sustain in 2026: patch software dependencies as CVEs emerge, monitor advisories for every piece of infrastructure they run, review logs and alerts from detection systems, conduct access reviews, manage certificates before they expire, audit third-party integrations, run penetration tests periodically, respond to incidents, maintain runbooks, train staff, and handle the compliance overhead that customers or regulators require. None of these tasks have an endpoint. Each one recurs on its own schedule, and the schedule is set by external parties.
The attacker’s work is bounded by success. They need one working path through one organization’s defenses, at some point in time. The target surface is large, changes constantly, and every new dependency, every new SaaS integration, every new API endpoint is a potential opening.
The xz-utils backdoor discovered in March 2024 illustrates this asymmetry with uncommon clarity. A sophisticated attacker, operating as the GitHub persona “Jia Tan,” spent roughly two years making legitimate, high-quality contributions to the xz compression library before inserting a backdoor into the build system. The payload hooked into OpenSSH’s RSA key decryption path through a chain of indirection that made it nearly invisible to automated scanning. It was caught because a Microsoft engineer noticed anomalous CPU usage and SSH latency on a Debian unstable system and followed the thread obsessively.
The total attacker investment was bounded: once the backdoor was merged, the work was done. The defensive work that followed is distributed across every organization that ships Linux, every maintainer who now thinks more carefully about who gets commit access, every security team that updated its threat model to account for multi-year social engineering campaigns targeting open source projects. That cost propagates indefinitely.
AI and the Attacker’s Cost Curve
The asymmetry between attacker and defender economics has always existed, but the AI tooling available now has sharpened it considerably.
LLMs reduce the marginal cost of producing convincing phishing content to near zero. Where a targeted spear-phishing email once required a human operator to research the target, draft a plausible message, and iterate, the same output now takes seconds of model inference. Research from ETH Zurich and several security vendors has shown that LLM-assisted phishing approaches the effectiveness of human-crafted attacks at a fraction of the cost. Voice cloning has extended this to vishing, enabling real-time impersonation of executives for business email compromise at scale.
Defenders get AI tools too: anomaly detection, automated triage, code scanning. The tooling is genuinely useful. But the defender’s position is structurally harder regardless of what tools are available, because the defender must block every attack across a target-rich environment while the attacker must find one working path through one organization. Better tools for both sides does not change that ratio.
The cost argument for attackers compounds further because AI-assisted attacks are composable. A phishing campaign can be automated, the reconnaissance feeding it can be automated, the resulting credential harvesting can feed an automated lateral movement tool. Each stage requires less human judgment than it did five years ago. The defender must respond to a system that scales without automatic scaling of their own response pipeline to match.
The Human Version of the Problem
The proof-of-work tax is not only computational. Security operations centers at large organizations generate hundreds of thousands of alerts per day, the majority of which are false positives. Industry surveys consistently find that SOC analysts spend 25 to 30 percent of their time triaging noise. The (ISC)² cybersecurity workforce study estimated a global gap of roughly 3.5 million unfilled security positions as of 2023. Burnout is the most commonly cited reason for attrition in SOC roles, and the pattern has been consistent across multiple years of workforce surveys.
This is the PoW problem in human terms. The system requires continuous human attention expenditure to function, and that expenditure has to come from somewhere. The pool of people willing to spend their careers watching dashboards and reviewing alerts is finite and under sustained pressure.
Larger organizations can staff this work. They have dedicated security teams, SIEM deployments, red teams, and bug bounties. Smaller organizations cannot. The ongoing work requirement functions as an effective minimum security budget floor; organizations below that floor are structurally exposed because the maintenance work required to close the gap is simply priced out of their reach. This explains something that raw breach statistics sometimes obscure: smaller organizations are disproportionately breached for structural reasons, not primarily architectural ones.
What Follows from the Framing
If the proof-of-work framing is accurate, a few things follow.
Approaches that reduce the surface area of ongoing work have more leverage than approaches that add tooling to manage the existing surface. Simpler architectures with fewer dependencies have less to patch. Managed services shift some of the PoW to vendors, introducing concentration risk in exchange for reduced per-organization overhead. The preference for minimal technology that the systems programming community often advocates for has a direct security implication: less code means less ongoing maintenance work, and less maintenance work means fewer lapses.
Regulatory pressure to push security costs upstream, toward the vendors who create complexity rather than the organizations that consume it, also fits this framing. If the PoW tax is systemic, changing where it falls changes the incentive structure. Secure by default becomes a supply-side requirement rather than a demand-side exercise.
The Hashcash model worked because it imposed costs proportional to attack volume on attackers while keeping legitimate use cheap. Modern organizational security defense inverts this: defenders pay costs proportional to their attack surface, which grows regardless of how carefully they build, while attackers pay costs proportional to their ambitions. That is not a problem that better tooling at the margin resolves. The proof-of-work framing makes visible why security budgets feel like they grow every year without the organization becoming meaningfully safer; the work requirement is structural, and seeing it clearly is at least a precondition for addressing it in ways that have a chance of working.