Ethereum Foundation opens an AI-agent bounty on the unproven math under zk proofs

The Ethereum Foundation went live today with better.codes, an open "autoresearch" challenge that asks solvers to point their own AI agents at a single cryptography problem and push a formally verified security bound higher, according to the Foundation's announcement on its blog (20 August 2026).

The target is koalaIRS12, a Reed–Solomon proximity problem lifted from the Foundation's Proximity Prize research and formalized end to end in ArkLib, the Lean 4 library for verified arguments of knowledge. Submissions are scored in bits: a comparator confirms an entry's exported theorem matches the pinned statement exactly, and the Lean kernel checks the proof. Accepted results are promoted to a public git repository, credited to both the solver and the AI model used. Solvers sign in with GitHub, clone the repo, and work inside a designated submission surface; the theorem statement, parameter point, and verification harness are all pinned. The Foundation says the challenge was built with Yukon and zkSecurity, though it doesn't describe what each contributed.

The stated reason matters more than the mechanism. Production hash-based SNARKs — including those securing zkrollups and zkVMs, and those the Foundation places at the center of Ethereum's post-quantum roadmap — depend on proximity gaps and correlated agreement for Reed–Solomon codes. Those systems target 128-bit security, and, in the Foundation's own words, that guarantee holds in full only if the underlying conjectures do. What's provable today stops short of what researchers believe is true. koalaIRS12 comes from Open Problems in List Decoding and Correlated Agreement by Gal Arnon, Dan Boneh and Giacomo Fenzi, the paper that framed the Proximity Prize grand challenges when the Foundation launched that initiative earlier this year.

Key facts

• Launched: 20 August 2026, at better.codes — Ethereum Foundation blog.
• Built by: EF Formal Verification team with Yukon and zkSecurity — EF blog.
• Problem: koalaIRS12, a Reed–Solomon proximity problem, formalized in ArkLib (Lean 4) — EF blog.
• Goal: raise the proven soundness lower bound to 128 bits; scored in bits, checked by the Lean kernel — EF blog.
• Lineage: drawn from the Proximity Prize, based on Arnon, Boneh and Fenzi's Open Problems in List Decoding and Correlated Agreement — EF blog.
• Prior challenges cited: ecdsa.fail, zk.golf, snark.fast — EF blog.

The real-world read

This is the Foundation's own launch post, so read the promotional parts as promotional. Three gaps stand out. First, no starting number: the post never states the current proven bound for koalaIRS12, which makes "raise it to 128" unmeasurable from the announcement alone — you cannot tell if the gap is two bits or fifty. Second, no money and no clock: awards and payments are deferred to unlinked "program terms" that the Foundation says "may be adjusted as the challenge progresses," and no deadline is given. Third, the claim that ecdsa.fail, zk.golf and snark.fast "have already moved research frontiers" is the Foundation's own assessment of its own programs, with no results cited.

The buried disclosure cuts the other way, and in the Foundation's favor: it is stating plainly that live systems handling real value rest on conjectures, not proofs. That is an unusually candid line for a launch post. The Foundation is also funder, problem-setter and evaluator here — the Lean kernel is the check on proofs, not on whether the chosen benchmark is the useful one.

Opinion, and whose

The thesis that "no single agentic setup is optimal across an open problem," so many independent setups beat one team, is the Ethereum Foundation's — presented as design rationale, not as a demonstrated result. The Foundation says it hopes to add further challenges over time; that is an intention, not a commitment.

Sources

• Ethereum Foundation blog, "Raising machine-checked security benchmarks to advance hash-based SNARKs through agentic collaboration," 20 August 2026 — primary source for the launch, mechanics, scoring, collaborators and stated goals. It is the Foundation's own announcement of its own program and carries a promotional frame; the factual mechanics are first-party, the claims about prior impact are self-assessed and uncorroborated.
• Arnon, Boneh and Fenzi, Open Problems in List Decoding and Correlated Agreement — cited by the Foundation as the origin of the koalaIRS12 problem and the Proximity Prize grand challenges. Referenced secondhand via the blog post.

Nothing here is financial advice.