Industries / Technology / Crypto and blockchain
Smart contracts
This is the work of writing and checking programs that run on a ledger and hold or move funds by fixed rules, such as trading pools, lending pools and bridges between ledgers.
5
Milestones achieved, from 1997 to 2022.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1997
Contracts written as code
In the September 1997 issue of First Monday, Nick Szabo wrote that "Smart contracts combine protocols with user interfaces to formalize and secure relationships over computer networks".
Credit
Nick Szabo
2015
A ledger that runs programs
On 30 July 2015 Ethereum went live with Frontier, "the first Live release of the Ethereum project", a ledger on which anyone can deploy programs that hold and move funds.
Credit
Ethereum Foundation and the project's developers
2016
A hack, then a rollback
The DAO, an investment fund run by smart contract, raised about 12 million ether (ETH) between 30 April and 28 May 2016. On 17 June an attacker used a flaw in its code to begin draining about 3.6 million ETH, a third of the total. On 20 July a majority of the network adopted a hard fork, a change in Ethereum's rules, that at block 1,920,000 moved the funds to a recovery contract, while a minority kept the old chain.
LessonUndoing a contract's result meant changing the ledger's rules for everyone, and not everyone agreed.
Credit
Ethereum's developers (fork); US SEC (record)
2017
A shared library deleted
In November 2017 a user triggered the self-destruct function of a shared library that Parity's multi-signature wallets relied on, freezing 513,774.16 ETH, about $169 million at the time. Parity said it had been warned of the flaw in August and planned to fix it "at some point in the future".
LessonMany contracts that call one shared piece of code share its single point of failure.
Credit
The Register (reporting Parity's post-mortem)
2022
A bridge emptied
On 29 March 2022 Ronin, the chain used by the game Axie Infinity, said 173,600 ETH and 25.5 million USDC had been drained from its bridge to Ethereum in two transactions. On 14 April the FBI said North Korea's Lazarus Group and APT38 were responsible for "the theft of $620 million in Ethereum reported on March 29".
LessonA bridge pools other people's tokens in one contract, which makes it one large target.
Credit
FBI (attribution); Silicon Republic (report)
Watch for
AI agents that exploit contracts
Anthropic researchers built SCONE-bench from 405 contracts exploited between 2020 and 2025; 10 AI models together wrote working exploits for 207 of them, worth $550.1 million in simulated stolen funds. On exploits that came after the models' training cutoffs they reached up to $4.6 million, and on 3 October 2025 two agents found two new flaws among 2,849 freshly deployed contracts, worth $3,694, all in simulators with no live funds touched.
Achieved whenan outside group reproduces the results on contracts and models of its own choosing.
Claimed by
Anthropic, December 2025
A benchmark for AI auditors
OpenAI and Paradigm released EVMbench, which scores AI agents on detecting, patching and exploiting smart-contract flaws taken from public audits, many from Code4rena contests, and from unreleased contracts. Paradigm says top models once exploited "less than 20%" of the critical bugs and that GPT-5.3-Codex now exploits "over 70%".
Achieved whengroups outside the two firms publish matching scores from the released tasks.
Claimed by
OpenAI and Paradigm, February 2026
Open
01
AI auditors scored in advance
A public test of AI audit tools on live audit contests, in which many reviewers hunt bugs in new code for prizes. Run each tool on a contest's code when the contest opens, lock its findings with a public time-stamp, and once the human results are judged, score the tool's precision and recall on high- and medium-severity bugs against them.
The briefCredit · unclaimed02
Clear-signing descriptions, checked
For the 100 most-used contracts on Ethereum, check whether a Clear Signing description (ERC-7730) exists, then test each one: run sample calls on a local copy of mainnet and compare what the description tells the user (action, token, amount, recipient) with what the call actually changes.
The briefCredit · unclaimed
Open 01 · the brief
AI auditors scored in advance
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of AI audit tools on live audit contests, in which many reviewers hunt bugs in new code for prizes. Run each tool on a contest's code when the contest opens, lock its findings with a public time-stamp, and once the human results are judged, score the tool's precision and recall on high- and medium-severity bugs against them.
Why it is still open
Today's tests look backward. Zealynx's AI Auditor Arena scores tools on 118 findings from 10 finished Code4rena contests, taking "the post-judging finding set as ground truth"; EVMbench draws on past audits; and Anthropic notes that because exploits are public, even its post-cutoff set "may still introduce some risk of data contamination". This research found no scoring done before results were known.
Built with what exists
Current AI models and coding agents that find software flaws (see the Note "Cyber defenders first"); contest code from Code4rena's public repositories; OpenTimestamps or a public Git commit to fix each set of findings in time; Foundry for proof-of-concept tests.
First step
Pick one contest opening this month, run one tool on its code on the first day, and publish the hash of its findings.
Done looks like
10 live contests, at least three AI tools, findings locked before judging, and precision and recall for each tool published beside the median human entrant.
Open 02 · the brief
Clear-signing descriptions, checked
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
For the 100 most-used contracts on Ethereum, check whether a Clear Signing description (ERC-7730) exists, then test each one: run sample calls on a local copy of mainnet and compare what the description tells the user (action, token, amount, recipient) with what the call actually changes.
Why it is still open
The Ethereum Foundation's launch says descriptions are checked "through independent reviews and attestations, and wallets decide which sources they trust", and gives no count of contracts covered. This research found no published check of the registry's accuracy.
Built with what exists
The registry and libraries at clearsigning.org; Foundry's anvil to copy mainnet locally, so no real funds move; Sourcify for verified contract source code; Google BigQuery's public Ethereum dataset to rank contracts by use.
First step
Take five descriptions from the registry and test 10 calls each on a local copy of mainnet.
Done looks like
100 contracts, the share with a description, 10 simulated calls for each described contract with every mismatch listed, issues filed with the registry, and a rerun each quarter.