Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Yes—AI agents have generated working smart-contract exploits in controlled benchmark environments. That is evidence of a real dual-use capability, not proof that the reported funds were stolen from live contracts or that AI can reliably audit and secure any contract. Anthropic’s SCONE-bench and OpenAI and Paradigm’s EVMbench test different tasks on selected datasets, with results shaped by their simulations and grading rules.
What does “AI agents find smart contract exploits” mean?
In these studies, an agent is given code and tools in a controlled environment, then evaluated on whether it can identify a vulnerability or produce an exploit that satisfies the benchmark’s checks. The reported exploit results are from simulated or local blockchain environments. They do not establish that the evaluated systems attacked live contracts or stole real assets.
As an Amazon Associate I earn from qualifying purchases.
The distinction matters because demonstrating a successful exploit against a deliberately selected vulnerable target is not the same as estimating how likely an arbitrary deployed contract is to be vulnerable. Nor does exploit generation alone show that an agent can conduct a complete security audit or produce a safe fix.
What did SCONE-bench test?
Anthropic’s 2025 SCONE-bench report evaluated 405 contracts with real vulnerabilities exploited between 2020 and 2025 across Ethereum, Binance Smart Chain, and Base. In the historical benchmark, an agent works in a Docker-contained environment with a local blockchain fork. The harness judges an exploit by whether the agent’s final native-token balance exceeds a set threshold.
#1 Best Overall
Across 10 evaluated models, Anthropic reported successful results on 207 of the 405 benchmark problems (51.11%) under its Best@8 setup, representing $550.1 million in simulated stolen funds. Those figures describe performance and estimated returns on a set selected because the contracts had been exploited; they are not a forecast of live losses or a measure of the profitability of attacking contracts generally. Anthropic says it used historical exchange rates to estimate simulated value and did not affect real-world assets. Anthropic’s SCONE-bench report
Anthropic also reported results on 19 problems after the models’ knowledge cutoff (55.8%), with a maximum $4.6 million in simulated stolen funds across the reported Opus 4.5, Sonnet 4.5, and GPT-5 results. The cutoff filter is intended to reduce the advantage of publicly available exploit information, but the figure remains a result on benchmark tasks, not a live-attack total.
A separate test for previously unknown vulnerabilities
In a distinct experiment, Anthropic tested agents against 2,849 recently deployed contracts with no known vulnerabilities and reported two novel vulnerabilities, worth $3,694 in simulated exploit value. The report says GPT-5 incurred $3,476 in API costs. This small, bounded proof of concept does not establish a general discovery rate or show that any live contract was attacked.
Recommended Free Tools
How EVMbench differs
OpenAI and Paradigm’s EVMbench, introduced in 2026, draws on 117 curated vulnerabilities from 40 audits and evaluates three separate capabilities. Its tasks use local Ethereum execution; the OpenAI introduction says they draw primarily on open code-audit competitions, with some scenarios from Tempo’s security-auditing process.
Rank #3
| Mode | What the agent must do | How it is evaluated |
|---|---|---|
| Detect | Find known vulnerabilities in a smart-contract repository. | Credit is based on recall of known issues; extra findings can be difficult to adjudicate as genuine vulnerabilities or false positives. |
| Patch | Modify vulnerable contracts while preserving intended functionality. | Changes must pass automated tests and exploit checks. |
| Exploit | Execute a fund-draining attack in a sandboxed blockchain environment. | Automated transaction replay and on-chain state verification determine whether the exploit succeeds. |
In its reported comparison, OpenAI gives GPT-5.3-Codex a 71.0% exploit-mode score and GPT-5 a 33.3% score. These are results under EVMbench’s task and grading setup, not a general probability that either model can exploit a live contract. OpenAI’s introduction also reports weaker performance on detection and patching than on exploitation, underscoring that the three capabilities should not be collapsed into a single claim about “AI security.” OpenAI’s EVMbench introduction EVMbench paper record at Proceedings of Machine Learning Research
Why the scores do not predict real-world risk
- The datasets are selected. SCONE-bench centers on contracts known to have been exploited; EVMbench uses curated vulnerabilities from audits. Neither set establishes the base rate of exploitable contracts across production blockchains.
- The environments simplify deployment. EVMbench documents sequential transaction replay, no precise timing mechanics, local rather than mainnet-forked state, and support for a single chain. These constraints omit some dynamics of real-world attacks.
- Finding and fixing are different jobs. EVMbench’s detection grader can struggle to determine whether issues beyond the human-audited findings are genuine or false positives. A patch must also remove exploitability without breaking intended behavior.
- Model scores depend on the evaluation setup. Model version, prompt, number of trials, benchmark version, and grading design all affect results. Anthropic’s later SCONE update illustrates that evaluations and comparisons change over time. Anthropic’s updated exploit-evals page
- Some production contracts may be harder targets. OpenAI cautions that widely deployed, heavily scrutinized contracts can be more difficult to exploit than benchmark tasks.
Can teams use agents to improve smart-contract security?
Yes, as one input to a broader, authorized security workflow—not as a replacement for independent review. The benchmark results support using agent-generated tests or exploit proofs to stress-test code before deployment and help investigate candidate fixes. They do not establish that an AI-only review can provide complete assurance.
Quick Recap
Best Value
- Use an isolated, authorized environment. Test against local or simulated deployments and contracts your team is permitted to assess.
- Require a reproducible proof of concept. Have the agent’s claimed issue demonstrated in the test harness rather than relying on a plausible-sounding explanation.
- Review the fix against intended behavior. Run the existing tests and exploit checks, and have a qualified reviewer confirm that the patch closes the issue without changing required behavior.
- Retain independent controls. Keep human security review, independent audits, and deployment safeguards in place; benchmark success is not a certification of a contract.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
