To Prompt Different...?
Disclaimer: I vibe thoughted this article
We’ve reached the point where large-scale vuln discovery and exploitation automation on open source code (so-called bug factories, if you wish) is the new baseline, a commodity, as predicted. At least, to some degree among security researchers - a fresh dev still wouldn’t have access or the perspective required, although they’d get it immediately after listing a bug bounty (and paying lots of money), hiring an SR (also paying, a bit less), or spending time learning the topic. In any case, the number of people to whom automation is a commodity is sufficient to make the distributed security market saturated at this new baseline level!
When I started working on Ethereum about half a year ago, every 4th report was a duplicate. Now, it’s getting closer to every 2nd or more, despite (or because of) reports becoming quicker to develop and more intricate (but paradoxically of the same severity as before). The BB is becoming more crowded and the client teams develop internal tooling (smart guys). I see similar stories on social media. BBs have traditionally been at the bleeding edge, the frontier. And when a technique becomes widespread enough, it naturally exhausts itself, setting a new level for manual, creative, craftful, and thorough research.
There’s another dimension to think about. Even for reports where I did use my expertise and collected experience to push the impact up from the baseline an agent can provide, it was still a dup because someone didn’t and sent raw slop as is, losing both my efforts and their potential.
A situation such as mine is likely ambiguous from the sponsor side too, as the info that provides higher severity normally won’t be rewarded and there’s now a moral and reputational embargo on shipping the new details (to warn users or develop a better fix, for example), as the SR may object (and rightfully so). If the sponsors don’t use the info, though, users may be harmed due to unawareness, causing even more damage. A good way to respond is for sponsors to issue smaller goodwill rewards if they want to utilize the value from the second report and improve their reputation too. Many don’t and prefer to free-ride on the unfortunate SR who invested their time, taking a gamble on karma (I hope there will be thought and popularized a new BB model, but that’s a topic for another time).
Therefore, I’m starting (as if I stopped!) to revise my workflows, because it’s becoming clear they need to get even more unique to find something no other people and agents can.
- For the past year, I saw an edge in management and full-cycle automation (read code -> collect findings -> test them -> prepare reports), as is the point of my writing, but it is now the baseline, commonly accessible to anyone who can be bothered enough. It is still a zero-to-one ability switch for most, especially in private engagements, but no longer a true difference maker.
- “Prompt different” like skills or other behavioral hints to agents don’t work that well anymore, evidently, because there are as many different prompts as there are SRs out there - and here I am, writing this text, + it’s now a common knowledge that skills of the kind “state every X before doing Y” often make SOTA models worse (re every article the big labs posted on thinning instructions and simplifying proceedures). Decomposing the thought process can produce something interesting, e.g., sending one agent to list every invariant, then have another work on that output. But that is most likely just my human mind tricking me, because as practice shows, such LLM gymnastics often cost more, work longer, and bring marginal improvements. They can be used to utilize safeguarded models within their policies and access greater intelligence at least for a part of the process, which is much more practical.
- “Use tools” (for discovery) on its own is also doubtful because models don’t yet know how to use them beneficially (re Dan Luu’s Astra testing writeup ), although this may also change as models develop - and become the new baseline.
Based on my vibes, the pursuit of a Critical has to make one step back, to the roots - manual reading and steering, balancing between autonomy and creativity, sacrificing the throughput for the sake of impact, amplifying efforts in the correct points where LLMs are potent, leaving other to humans.
Congrats if you never changed your process anyway, lol (I knew a few people), but this is not the way I like. I have troubles sitting still in one place. Every time a new model came out in the last year, I tried to push it all the way up the abstraction to make decisions on different levels. First little small choices, now it self-organizes and can run campaigns with very little supervision. The issue is, even when it can loop the full cycle from an idea to the report, however well or poorly it does, the quality of the ideas it can produce is subject to the thing everyone knows and writes about from time to time. Entropy.
Think, how many actively working security researchers do we have? Thousands to tens of thousands?
“My estimate for late 2026 is around 50,000–150,000 actively working security researchers worldwide” ChatGPT 6 xhigh
“My estimate is about 40,000–80,000 people actively finding and reporting vulnerabilities worldwide right now” Opus 5.5 high
Each of those ~100k human beings has a unique perspective when it comes to code, vulnerabilities, ideas. Take, let’s say, the top 5 among them, the absolute best ones, beasts who find bugs where no one else can. Imagine you can erase their memory. Hire for an engagement. And compare the results. Erase memory. Repeat 100 times (in vacuum, no washroom time). Ah, and they are quintuplets, born on the same day, to the same mother, then raised under one roof with home education. You get the idea. You will see mostly the same list. Those 5 guys today are GPT, Claude, DeepSeek, maybe Grok and GLM (not yet top five worldwide ofc, but have you seen news?)
Now imagine each dev out there does the same, asking the same SRs to review their code (and to write it too!). However you frame the task, you’ll see the same converging results if you don’t pump your own, human and, therefore, unique, perspective (yet another topic for an article - entropy and diversity).
So:
- Both prompts and tools can only provide unique results above the baseline when there’s a human to provide timely, thoughtful, and (ideally, but not necessarily) educated, input.
This is how I understand what’s happening in security right now.
I don’t think this is necessarily a bad state of things overall, as the role of security is to secure things, and, as I mentioned before, we’ve also got a baseline in code too. Agents write better code and discover more bugs than the majority of humans who were or are involved (fair to call superhuman? In most cases), as evident by the number of vulnerabilities disclosed in the past few weeks, all in old, 100% handmade, organic code.
As for my workflow, it’s similar to many that popped up in individual SRs’ home labs and even big AI labs (re @bebuu_ai’s post ), and that now collide in BBs. It is now a part of that greater good, just moving behind private paid engagements, as a better use of economic resources. Or it becomes the base for future developments that will never stop.
Thanks Alex Filippov for giving the idea for the article