Notes from the Ethereum Basis’s Protocol Safety group on working coordinated AI brokers in opposition to actual protocol code, together with how we manage the work, what holds up beneath scrutiny, and what shopper groups and safety researchers can take from it. This publish 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 group, we have been working coordinated AI brokers in opposition to the sorts of techniques the community will depend on, like techniques 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, mounted and disclosed as CVE-2026-34219 with credit score to the group.
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 actual bugs from those that simply appeared actual.
This publish is for shopper groups and safety researchers who wish to do the identical factor. It covers how we manage the brokers, the bar a candidate has to clear earlier than it counts as a discovering, and the habits that preserve the outcomes reliable.
Groups elsewhere are converging on the identical recipe. Anthropic’s Frontier Crimson Group constructed an agent that writes property-based checks and located actual bugs throughout the Python ecosystem. Cloudflare ran a frontier mannequin via a security-research harness in opposition to their very own techniques. Everybody lands on the identical loop: level a succesful mannequin at a codebase, let it search, and triage what comes again. So the actual 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 old-fashioned in just a few weeks. So this publish is intentionally concerning the strategies, that are persistent, slightly than the tooling. Disclosure is its personal matter and can most likely be its personal publish.
An agent pointed at a codebase is a search software, quite a bit like a fuzzer. The distinction is what comes again. A fuzzer fingers you a crash and a stack hint. An agent fingers you much more, together with a write-up (name chain, affect declare, prompt severity) and the artifacts to again it, like a proof-of-concept you may run in opposition to the actual code.
All of that makes the consequence straightforward to learn and simple 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 become actual.
How the work is organized
We run many brokers in parallel in opposition to one goal. They coordinate via 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 received this method from Anthropic’s writeup on constructing a C compiler with a fleet of brokers, which coordinates the identical means. There is no central coordinator to construct or preserve, and fewer that may go fallacious.
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 preserve, the way in which it would break, and the proof that will settle it.”
- Searching takes one speculation, traces the code path, and tries to construct a reproducer.
- Hole-filling appears to be like at what was accepted and what was rejected, writes the following batch of hypotheses, and tracks protection so the brokers do not preserve 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 looking, unbiased validation, deduplication, reporting, and their writeup helped form ours.
This is what a candidate appears to be like like earlier than it counts as a discovering:
goal: element and entry level an attacker can really attain invariant: the property that should maintain mechanism: the precise means 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 actual 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 accomplished. An agent that has to put in writing down an observable proof cannot fall again on “this appears to be like dangerous.”
Reproducible or it did not occur
One rule issues greater than another. A candidate is not a discovering till there is a self-contained artifact that reproduces the failure in opposition to the actual code, and that runs for somebody who did not write it.
The reproducer does not learn the write-up, and it does not care how assured the mannequin sounded. It both runs or it does not.
Most of its worth is within the false positives it catches. Three of them come up again and again, and every one is the agent getting a go for the fallacious cause:
- A panic that solely occurs in a debug construct. Compile and run it the way in which the software program really ships, and the worth simply wraps round. Nothing crashes. It appears to be like 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 means.
- In formal-verification work, a proof that goes via however doesn’t suggest what you needed. 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 happy, however the theorem does not constrain the conduct you really cared about.
None of that is new. It is the identical factor as a check that passes as a result of it does not really examine something. What’s new is the quantity. An agent writes the ineffective model as quick as the actual one, and simply as confidently. So the examine must be computerized. You may’t rely on the agent to catch itself.
Sign-to-noise is a lot of the work
Most candidates are fallacious, duplicate, or out of scope. That is not an issue with the strategy; that is the way it works. The purpose is to reject the fallacious ones quick and again the actual ones with proof that is arduous to argue with.
Each candidate that survives will get two unbiased checks. Can an actual attacker really attain it in a standard configuration? And what does it value the attacker to drag off, in comparison with what it prices the community if it really works? A bug that any single peer can set off could be very totally different from one which wants particular entry or an enormous quantity of sources.
All the pieces will get checked in opposition to a working record of what is already recognized, mounted, or rejected. With out that, the brokers preserve rediscovering the identical closed challenge and reporting it repeatedly.
Acceptance charges range 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 continues to be value understanding. “We appeared arduous and located nothing” is an actual consequence. 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 via.
We’re not the one ones who discovered that the triage is the arduous half. Cloudflare’s major 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 professional assessment to get right down to a prime tier that held up about 86 p.c of the time. The era was the straightforward 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 methodology.
What the brokers are good at, and the place they mislead
There’s hype in each instructions, so here is a plain record of what the brokers do properly and the place they mislead.
| Good at | Deceptive at |
|---|---|
| Studying the spec and the code collectively | Name chains that look reachable however aren’t |
| Stating and checking an actual invariant | Gaming the success examine (a go for the fallacious cause). |
| Drafting a reproducer from a one-line concept | Inflating severity to match how dramatic the write-up sounds |
| Suggesting a root trigger earlier than you have appeared | Bugs that span a sequence of legitimate steps |
The cut up is not even regular from one process to the following. 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 may’t assume one good consequence means the following 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 fallacious. For these, the agent is not the search software. Its job is to recommend which sequences are value working via a stateful check harness. Used that means, it really works properly. Used as a substitute for the harness, it misses the costliest bugs there are, those that solely present up throughout a sequence.
Retaining it trustworthy
A number of habits do a lot of the work of constructing 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 needs to be one thing you may re-run months later.
- Determinism the place it counts: one setting, one technique 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 means over-specified checks do, they preserve following the steps after the steps cease making sense. A examine of repository context recordsdata discovered the identical factor: the additional necessities lowered process success and raised value by over 20%, and the authors suggest maintaining context to the minimal necessities.
- An individual makes the ultimate name: brokers recommend. They do not resolve what’s actual, what’s a replica of a recognized challenge, or what will get disclosed and when.
The bottleneck moved
AI did not change 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, maintaining the record of recognized 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 really issues. But it surely’s nonetheless a bottleneck, and ignoring that’s how you find yourself delivery a fallacious “it is wonderful.”
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 commonplace observe over the past fifteen years. The instruments are new. The practices aren’t.
How briskly the instruments preserve altering is an open query. Nicholas Carlini, cautious and as soon as a skeptic himself, argues the exponential case is value taking critically, even whereas he retains extensive error bars on it. If the era facet climbs that quick, the judgment facet has to climb with it, or the hole between what will get produced and what really will get verified solely widens.
For the techniques Ethereum will depend on, that is the half that issues. Brokers allow us to cowl way more floor than we might by hand. In alternate, they ask for extra cautious judgment, throughout a a lot greater pile of confident-sounding claims. That is a commerce value making, so long as you keep in mind that the judgment is the actual product.