Notes from the Ethereum Basis’s Protocol Safety crew on working coordinated AI brokers in opposition to actual protocol code, together with how we set up the work, what holds up beneath scrutiny, and what consumer groups and safety researchers can take from it. This submit stands by itself; later posts will go deeper on particular person purchasers.
What we have been working, and what stunned us
On the Ethereum Basis’s Protocol Safety crew, we have been working coordinated AI brokers in opposition to the sorts of methods the community is dependent upon, like methods software program, cryptographic code, and contracts that should be proper. The brokers discovered actual bugs. One is now public: a remotely-triggerable panic in libp2p’s gossipsub, a core a part of the peer-to-peer layer Ethereum consensus purchasers run on, fastened and disclosed as CVE-2026-34219 with credit score to the crew.
Brokers discovering bugs wasn’t the shock. The shock was how little of the work went into discovering them, and the way a lot went into telling the true bugs from those that simply appeared actual.
This submit is for consumer groups and safety researchers who need to do the identical factor. It covers how we set up the brokers, the bar a candidate has to clear earlier than it counts as a discovering, and the habits that hold the outcomes reliable.
Groups elsewhere are converging on the identical recipe. Anthropic’s Frontier Purple Crew constructed an agent that writes property-based assessments and located actual bugs throughout the Python ecosystem. Cloudflare ran a frontier mannequin by a security-research harness in opposition to their very own methods. Everybody lands on the identical loop: level a succesful mannequin at a codebase, let it search, and triage what comes again. So the true query is how to do that with out drowning in confident-sounding noise.
One caveat up entrance: tooling for agent-driven audits strikes quick, and any particular setup is outdated in a couple of weeks. So this submit is intentionally concerning the strategies, that are persistent, fairly than the tooling. Disclosure is its personal matter and can most likely be its personal submit.
An agent pointed at a codebase is a search software, quite a bit like a fuzzer. The distinction is what comes again. A fuzzer palms you a crash and a stack hint. An agent palms you much more, together with a write-up (name chain, impression declare, advised severity) and the artifacts to again it, like a proof-of-concept you’ll be able to run in opposition to the true code.
All of that makes the outcome straightforward to learn and straightforward to belief, the working proof-of-concept most of all. So do not rely what number of candidates an agent produces. Rely what number of transform actual.
How the work is organized
We run many brokers in parallel in opposition to one goal. They coordinate by the repository itself, with shared state in model management and no central course of handing out work. An agent writes down a declare the place the others can see it, does the work, and commits.
We acquired this method from Anthropic’s writeup on constructing a C compiler with a fleet of brokers, which coordinates the identical method. There is no central coordinator to construct or keep, and fewer that may go incorrect.
The roles are generated by the work that is found:
Recon turns an assault floor into concrete, testable hypotheses. Not “audit the decoder” however “this subject is trusted previous this level; here is the property it ought to hold, the best way it’d break, and the proof that may settle it.”Searching takes one speculation, traces the code path, and tries to construct a reproducer.Hole-filling appears at what was accepted and what was rejected, writes the subsequent batch of hypotheses, and tracks protection so the brokers do not hold going over the identical floor.Validation re-checks every candidate independently, removes duplicates, and decides.
We did not invent this pipeline. Cloudflare describes the identical levels, recon, parallel searching, impartial validation, deduplication, reporting, and their writeup helped form ours.
This is what a candidate appears like earlier than it counts as a discovering:
invariant: the property that should maintain
mechanism: the particular method it is perhaps made to break
success: observable proof: a panic, a stall, an accepted-invalid enter
reproducer: a self-contained artifact that runs in opposition to the true code
dedup: a key, so two brokers do not chase the identical factor
The schema is there for a cause. It forces a particular, testable declare and a transparent definition of achieved. An agent that has to put in writing down an observable proof cannot fall again on “this appears dangerous.”
Reproducible or it did not occur
One rule issues greater than every other. A candidate is not a discovering till there is a self-contained artifact that reproduces the failure in opposition to the true code, and that runs for somebody who did not write it.
The reproducer would not learn the write-up, and it would not care how assured the mannequin sounded. It both runs or it would not.
Most of its worth is within the false positives it catches. Three of them come up time and again, and every one is the agent getting a go for the incorrect cause:
A panic that solely occurs in a debug construct. Compile and run it the best way the software program truly ships, and the worth simply wraps round. Nothing crashes. It appears like a crash, however it is not one.A reproducer that builds some inner worth by hand, one no actual enter might ever produce, as a result of each path an attacker controls rejects it earlier. The bug solely “reproduces” in opposition to a perform that nothing reachable calls that method.In formal-verification work, a proof that goes by however doesn’t suggest what you wished. The assertion is trivially true no matter what the code does, or it is weaker than the property you meant to seize. The verifier is glad, however the theorem would not constrain the conduct you truly cared about.
None of that is new. It is the identical factor as a take a look at that passes as a result of it would not truly test something. What’s new is the quantity. An agent writes the ineffective model as quick as the true one, and simply as confidently. So the test needs to be automated. You possibly can’t rely on the agent to catch itself.
Sign-to-noise is a lot of the work
Most candidates are incorrect, duplicate, or out of scope. That is not an issue with the strategy; that is the way it works. The aim is to reject the incorrect ones quick and again the true ones with proof that is onerous to argue with.
Each candidate that survives will get two impartial checks. Can an actual attacker truly attain it in a standard configuration? And what does it value the attacker to tug off, in comparison with what it prices the community if it really works? A bug that any single peer can set off may be very completely different from one which wants particular entry or an enormous quantity of assets.
Every part will get checked in opposition to a working listing of what is already identified, fastened, or rejected. With out that, the brokers hold rediscovering the identical closed problem and reporting it repeatedly.
Acceptance charges differ quite a bit from goal to focus on, and that variation is beneficial by itself. Run this in opposition to mature, closely audited code and virtually nothing survives, which remains to be price figuring out. “We appeared onerous and located nothing” is an actual outcome. Run it in opposition to less-explored code, or in opposition to formally verified code, the place a machine-checked proof covers a mannequin and the deployed bytecode is just assumed to match it, and extra will get by.
We’re not the one ones who discovered that the triage is the onerous half. Cloudflare’s principal takeaway was {that a} slender scope beats broad scanning. Anthropic’s property-based-testing agent generated one thing like a thousand candidate experiences, then used rating and skilled assessment to get right down to a high tier that held up about 86 p.c of the time. The technology was the simple half. I am not going to publish our personal numbers right here; tied to a particular goal, they’d say extra concerning the goal than concerning the technique.
What the brokers are good at, and the place they mislead
There’s hype in each instructions, so here is a plain listing of what the brokers do effectively and the place they mislead.
The break up is not even regular from one process to the subsequent. Stanislav Fort, testing a spread of fashions on actual vulnerabilities, calls this a jagged frontier, or a mannequin that recovers a full exploit chain on one codebase can fail fundamental data-flow tracing on one other. You possibly can’t assume one good outcome means the subsequent will maintain up, which is another excuse each candidate will get checked by itself.
The final row is the vital one. A single agent session is sweet at one-shot reasoning and unhealthy at bugs that span a sequence of steps, the place every step is legitimate and solely the order is incorrect. For these, the agent is not the search software. Its job is to counsel which sequences are price working by a stateful take a look at harness. Used that method, it really works effectively. Used as a alternative for the harness, it misses the most costly bugs there are, those that solely present up throughout a sequence.
Holding it trustworthy
A couple of habits do a lot of the work of creating agent findings reliable, and none of them are difficult.
Provenance on each artifact: what produced it, with what context, in opposition to which revision. A discovering ought to be one thing you’ll be able to re-run months later.Determinism the place it counts: one setting, one method to construct and run, so “reproduces” means the identical factor on each machine, not simply the one the place it was discovered.Norms, not scripts: inform brokers what issues, the invariants and the bar for an actual discovering, as a substitute of a numbered process. Over-scripted brokers break the identical method over-specified assessments do, they hold following the steps after the steps cease making sense. A research of repository context recordsdata discovered the identical factor: the additional necessities lowered process success and raised value by over 20%, and the authors advocate retaining context to the minimal necessities.An individual makes the ultimate name: brokers counsel. They do not resolve what’s actual, what’s a reproduction of a identified problem, or what will get disclosed and when.
The bottleneck moved
AI did not exchange the safety researcher. It moved the work. The time that used to enter arising with and chasing down hypotheses now goes into judging them at scale, together with constructing the oracle, working the triage, retaining the listing of identified points, and dealing with disclosure.
The bottleneck did not go away. It moved from discovering bugs to trusting the outcomes, which is a greater place for it, as a result of that is the place human judgment truly issues. Nevertheless it’s nonetheless a bottleneck, and ignoring that’s how you find yourself delivery a incorrect “it is tremendous.”
The practices that make this work aren’t new. Reproducible failures, actual oracles, and cautious triage are the identical practices that turned fuzzing from a analysis matter into normal follow during the last fifteen years. The instruments are new. The practices aren’t.
How briskly the instruments hold altering is an open query. Nicholas Carlini, cautious and as soon as a skeptic himself, argues the exponential case is price taking significantly, even whereas he retains huge error bars on it. If the technology aspect climbs that quick, the judgment aspect has to climb with it, or the hole between what will get produced and what truly will get verified solely widens.
For the methods Ethereum is dependent upon, that is the half that issues. Brokers allow us to cowl way more floor than we might by hand. In trade, they ask for extra cautious judgment, throughout a a lot greater pile of confident-sounding claims. That is a commerce price making, so long as you keep in mind that the judgment is the true product.
