Are Security Researchers Doomed? History Says No

At DEF CON this year, a senior security researcher at a FAANG company told a junior colleague something blunt: her job is basically useless because of AI. Security researchers, he said, are doomed. The junior researcher, Garance, took the question to her boss — Roni Carta, better known online as Lupin, founder and CEO of the supply-chain security firm Depi and creator of the npx Confusion vulnerability class.

What followed was not a comfort piece. Instead of telling her "don't worry, everyone survives," Carta walked through five centuries of technological history — printing presses, Luddites, photography, electricity, computers, chess — and the history of vulnerability research itself. The answer he landed on is uncomfortable in the way reality usually is: the field survives and grows, but survival is conditional on moving to where the new value is.

The oldest mistake in the book

Whenever a machine learns to do something we previously associated with human skill, somebody always makes the same leap: the machine can do the task, therefore the human who does the task is doomed. We have been making that jump for centuries.

In 1492, the German abbot Johannes Trithemius wrote In Praise of Scribes, worried about monks whose job was copying books by hand. His arguments sound eerily modern: printed books used paper instead of durable parchment; the act of copying itself had value — a monk reading, memorising and learning while he worked; a machine removes the process through which people actually get good at a craft. You have heard all three arguments applied to AI. I have probably made all three myself.

Trithemius was not stupid. That is the point. He could correctly identify the things a printed book did worse than a manuscript and still completely miss what happened next. Because the winning question was never "is a printed book better than the best manuscript?" It was: what happens when making another copy stops requiring months of human labour? Once you ask that question, you are talking about a different world.

The uncomfortable part: sometimes the scared people were right

History does not let us off with a "technology always creates more jobs" shrug. The Luddites were not cartoon technophobes. Many were highly skilled textile workers, and they were not wrong about what was happening to them. The shearing frame did the work of a skilled cropper — two unskilled workers with a machine could do in a day what a skilled man did in a week. Nobody had to be out-crafted.

The model does not need to become the best researcher in the world at a week-long source-code audit. If a company needs ten people for a class of work and AI lets three do it, seven people still have a problem. Even David Ricardo, the economist, revised his position in 1821: machinery could genuinely hurt workers' interests. The workers who testified before the Sadler Committee in 1832 described exactly what they could see — one man thrown out of work while another was pushed to run machinery day and night.

Photography, electricity, computers: the pattern repeats

Every generation produced extremely competent people explaining exactly why the new technology sucked. Charles Baudelaire attacked photography in 1859 with what are now AI-tweet arguments: "these people aren't real artists," "the machine is just reproducing reality," "people are confusing technical output with actual art." Photography did automate part of what painters were paid for. And painting did not disappear — it moved to what the camera could not give you. Producing a realistic image and being an artist are not the same thing.

Henry Morton called Edison's lamp a "conspicuous failure." William Preece, a leading electrical engineer, produced a full analysis proving that lighting a house with several bulbs was practically impossible. They understood the technology better than anyone. They were still wrong, because the technology was improving fast. "This is a real limitation" and "this limitation protects my profession" are two completely different claims. The first is engineering. The second is a bet on the future.

Computers and the internet did not teach us anything either. Paul Krugman predicted in 1998 that the internet's economic impact would rival the fax machine. Robert Metcalfe, the Ethernet inventor — about as inside as you can get — predicted the internet would catastrophically collapse in 1996 and ended up eating a blended copy of his own column on stage. But computers also show the other side: newspaper typographers were genuinely hit, and their union leader Bertram Powers negotiated a brutal, brilliant deal — newspapers could automate, but existing members got lifetime employment guarantees while the trade itself was allowed to slowly disappear. He stopped arguing with the existence of the machine and started asking what position humans could still defend once the machine existed.

The closest parallel: chess

Chess was treated as the purest evidence of human intelligence until Deep Blue beat Kasparov in 1997. The central task was automated, and the best human alive lost. Yet humans did not stop playing chess. The machine became part of the game — players trained with engines, analysis got vastly deeper, decades-old opening theory was re-evaluated. Kasparov himself became the champion of human-machine collaboration, and in 2005 a pair of amateurs running ordinary chess programs beat grandmasters with supercomputers. The value moved, then moved again. That is the pattern hiding under all these stories: we are terrible at separating a task from the larger reason we valued the human doing the task in the first place.

Vulnerability research has already been declared dead four times

Here is the part that matters most for people in security, because it is recent and in-domain. Vulnerability research has been declared dead at least four times:

1. Fuzzing, 1990. Barton Miller crashed a quarter to a third of standard UNIX utilities with random input. Reaction: bug hunting is now a button press, the craft is over. What happened: fuzzing became table stakes, trivial bugs got mined out, and researchers moved up into grammar-aware fuzzing, coverage guidance, sanitizers and reachability analysis. The tool removed the easiest tier and created three harder ones above it.
2. Metasploit, 2003–2007. Anyone can exploit now, so exploit development is dead. Instead it moved into mitigation bypass — ASLR, DEP, stack cookies, CFI. The bar went up, and the population that could clear it went up too, because the tooling made the entry ramp survivable.
3. DARPA Cyber Grand Challenge, 2016. Seven machines played autonomous capture-the-flag with no humans, in the same building where DEF CON was opening. Headlines screamed that hackers no longer needed to be human. The winning system, Mayhem, then entered the human DEF CON CTF the same week and finished last — exactly as DARPA's own program manager had predicted. The revolution was tooling, not replacement.
4. The recon and tooling era, 2015–now. Subdomain enumeration, nuclei templates, mass ASN sweeps. "Recon is commoditised, the bounty hunter is finished." Raw recon became worthless precisely because it was automated, and the value moved to target selection, understanding trust boundaries nobody documented, and knowing which of ten thousand findings actually matters.

Ten years after the Cyber Grand Challenge, DEF CON is bigger, and vulnerability research is a larger and better paid field than it was in 2016. Every single time, the automation ate the bottom of the stack and the humans moved up.

What automation does not eat

The claim is not "everyone survives." The croppers did not survive. The handloom weavers did not survive. The typographers did not survive as a trade. If your entire professional value is "I can find a memory-safety bug in a parser faster than average," be worried — on a timescale of a couple of years, not decades.

What gets automated is execution of a known technique against a known target class. What does not get automated, so far, is deciding what to look at, understanding a trust boundary nobody wrote down, and inventing a vulnerability class that did not exist yesterday.

Carta uses his own work as the cleanest illustration. At DEF CON 33 he named npx Confusion — a dependency-confusion variant where an unscoped binary referenced in package.json's bin/scripts can be claimed on the public npm registry by anyone. His team demonstrated it by backdooring Fortune 500 companies' CI pipelines and pocketing over $50K in bounties; a follow-up scan found 128 unclaimed "phantom" packages downloaded 121,539 times in seven months. Since then, a third-party tool called npxconfuse has automated detection of exactly what he described.

His reaction is the thesis in one sentence: the class was the contribution. Automating its detection is the proof the contribution was real — and it means the next contribution has to be a class nobody has named yet. You get automated from behind, not ahead. The tools eat your published work, which is what publishing is for, and you go find the next thing.

The two kinds of wrong

History has two groups of people. The people who said "the machine won't work" were wrong every single time — Morton on the light bulb, Preece on subdivision, Krugman on the internet, Metcalfe on a machine that was already working. The people who said "this will cost us something real, and somebody should govern the transition" were right every single time — Ricardo, Drake, the Luddites, Powers.

When someone tells a security researcher her job is useless because of AI, they are making the first argument, inverted: a confident capability prediction about a fast-moving technology, on no evidence, in a domain they do not work in, to someone who does. Morton did that. Preece did that. They had better credentials than most people making that claim today, and they were still wrong.

And there is a second, more common possibility: some people saying it is over are not scared. They just cannot see where the new value went. That is a harder failure to admit than fear, because fear at least means you are paying attention.

The actual advice

Your job is not safe. Nobody's job has ever been safe, and anyone who promises you otherwise is selling something. What is safe, on a two-hundred-year record with no real counterexample, is the position: the person who finds what the current tooling cannot see yet. That position keeps existing. It just keeps moving.

The tools do not move you. You move. Powers put on a disguise, walked into a printing trade show, and looked at the machine that was coming for his members with his own eyes. It is the single smartest thing anyone in this entire story did. Go look at the machine.

Why this matters for a developer blog

This is not abstract philosophy for security teams — it is the same pattern hitting software development and content work right now. The human review step in AI writing pipelines exists for exactly this reason: the model can produce the output, but deciding what is correct is still a human judgement. And agentic AI is already reshaping how scientific software gets maintained — verification and stewardship, not implementation, becoming the bottleneck.

The tools keep coming. The bottom of the stack keeps getting automated. And the people who keep finding what the current tooling cannot see yet are the ones who stay valuable. The machine did not kill the scribes by being a better scribe. It killed them by making a copy cost nothing. Learn the difference, and go look at the machine.

Read the original thread on X: 0xLupin — "security researchers are doomed".