Argus

Eth Security is an independent non-profit organization securing the Ethereum network using novel methods. Today, conventional security in the Ethereum ecosystem is built around isolated snapshots: pre-deployment audits examine one repository at a fixed point in time, and bug bounties cover only what individual projects can scope and fund. Yet Ethereum in production is a tightly coupled stack whose codebases change continuously, where a flaw in a client, prover, or shared library can put downstream protocols and user funds at risk even as AI lowers the cost of offensive discovery. Defending a live network under those conditions requires continuous, ecosystem-wide scanning rather than one-off review.

To meet that need, Eth Security operates Argus, a continuous, self-improving bug-hunting engine that points frontier AI at the code the Ethereum ecosystem depends on, and el-bigote, its working reference implementation: the orchestration, methodology, and knowledge layer that runs the full find-to-confirm loop today, unified with a local-model workhorse that handles the high-volume code reading on owned hardware. It runs now as a live system, and its proof is the body of real vulnerabilities it has already found.

Together, Argus and el-bigote aim at the whole Ethereum dependency graph rather than any single bug-bounty program's scope: the base-layer execution and consensus clients and core-protocol code, the L2 rollups and their zero-knowledge proving systems, restaking and data-availability infrastructure, bridges and cross-chain interoperability, stablecoins, the major DEX, lending, and perpetuals protocols, wallets and custody, and account-abstraction and MEV infrastructure. That is several hundred distinct targets across more than a dozen categories, including the highest-consequence base-layer code and the shared dependencies that no single bounty is funded to review.

Across that map the system has surfaced a large body of high- and critical-severity issues, in live mainnet protocols and in the base-layer client and consensus code beneath them. The finds concentrate in the hard classes that separate real capability from shallow pattern-matching: consensus and liveness halts, unmetered and quadratic-work denial of service, cross-margin and settlement accounting errors, bridge value-conservation gaps, and zero-knowledge soundness bugs such as under-constrained witnesses in rollup provers, where missing constraints let invalid state transitions pass verification. The same engine that finds a chain-halt in a perpetuals appchain also finds a resource-exhaustion path in a consensus client and a soundness gap in a ZK circuit, and that reach is the signal of a single general-purpose technology rather than a narrow point tool. A meaningful share of these findings independently rediscovered issues that top human auditors had already confirmed as real and high-severity, the clearest evidence there is that the engine finds what the best researchers find, and alongside them it has turned up genuinely novel issues. In one representative case on a major perpetuals protocol, this local-to-frontier funnel rediscovered a confirmed high-severity bug and surfaced a novel divide-by-zero while the frontier model read only a handful of functions instead of an entire module, the kind of economics that lets the technology run at ecosystem scale.

The capability is a property of the whole pipeline. Argus routes each task to whichever model does it best at the right cost, measured by recall per task type; it generates leads across many diverse analysis lenses, refutes false-positive noise automatically, and confirms only what survives with reachability analysis, differential testing, and fuzzing, so nothing reaches a maintainer on a model's say-so. It treats coverage as perishable, pinning every pass to an exact commit, re-scanning as new code lands and as old audits age, and folding fresh disclosures and research into new lenses that sweep every target the pattern fits. Every dead end is remembered so it is never re-derived, and every prompt that works is generalized across every fork and cousin that shares the pattern, so each pass costs less and reaches deeper than the last.

Behind the technology is a team with a proven, hands-on track record of finding high-profile bugs in the Ethereum ecosystem. Eth Security is founded and led by David Theodore, whose background brings together three disciplines that rarely sit in one person: exploit development for the US national security community, security engineering at Google, and protocol security from inside the Ethereum Foundation as a founding member of its Protocol Security Team. The AI supplies scale and tireless breadth; senior researchers set the aim, verify and file every finding through coordinated disclosure, and keep the final word, under a defense-only charter.