Breaking

Ethereum Foundation Uses AI Agents to Hunt Protocol Bugs; Most Flags Are False Positives

July 9, 2026 at 23:22 EDT

AI Agents
Security
Open Source

Ethereum Foundation's protocol security team deployed multiple AI agents against Ethereum's core code to hunt for bugs, and on July 9, 2026 disclosed both the results and the challenges. The agents found real vulnerabilities, but the vast majority of flagged candidates turned out to be false positives, with sorting the real from the fake (triage) becoming the bulk of the work.

July 9, 2026 · Ethereum Foundation Protocol Security

AI agents found real Ethereum bugs — but sorting the fakes became the real job

Multiple coordinated AI agents hunted through Ethereum's core code and surfaced genuine vulnerabilities. Yet the vast majority of flagged candidates were false positives — turning triage, not discovery, into the product.

CVE-2026-34219

Real remotely-triggerable crash found in libp2p's gossipsub — published, fixed, disclosed

Mandatory fields a candidate must satisfy before it counts as a finding

Central coordinators — agents share state via the repository (fleet-of-agents)

Where the work actually goes

Finding candidates takes little effort. The bulk of the labor is triage — separating genuine bugs from the many that only look real.

Small

Search
(generating candidates)

Bulk

Triage
(sorting real from fake)

Why candidates fail

Panics that only occur in debug builds
Reproductions relying on inputs an attacker cannot reach
Trivial or meaningless formal proofs
Misjudged call-chain reachability
Overestimated severity

A valid finding must specify

Target — an attacker-reachable component
Invariant — a property that must hold
Mechanism — how it breaks
Proof — observable evidence
Reproducer — runs in real code
Dedup key — prevents duplicates

“A search tool, not an oracle.”

Agents excel at one-shot reasoning but stumble on multi-step bugs where each step is valid and only the ordering is wrong. Independent human verification and deduplication remain essential — the methodology is reproducible proof plus final human judgment, not any single product.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.