AI smart contract security is moving beyond pattern-based scans toward agents that can test contracts, use dependencies and simulate attacks. In a November 19 report, U.Today reviewed researc
AI smart contract security is moving beyond pattern-based scans toward agents that can test contracts, use dependencies and simulate attacks. In a November 19 report, U.Today reviewed research showing agents can reproduce known exploits and uncover some new flaws. Most results come from controlled experiments, not evidence of AI-driven mass theft.
Key Takeaways
- Anthropic agents reproduced exploits against 207 of 405 historically vulnerable contracts, representing $550.1 million in simulated stolen funds, not real losses.
- Benchmarks show uneven performance in finding, exploiting and patching flaws.
- Reported crypto theft figures don’t establish that AI caused the attacks.
- Rigorous testing and monitoring could help teams respond to security risks.
What happened in AI smart contract security
U.Today reports that AI agents can do more than search code for familiar vulnerability patterns. They can combine contract code with dependencies, a local blockchain fork and decompilation, then test how the pieces interact. This may help probe access-control errors, accounting flaws, oracle weaknesses and unexpected contract behavior.
Keeping source code private may offer less protection if attackers can inspect deployed bytecode and test it. But a working exploit in a sandbox isn’t the same as taking funds from a live protocol.
Anthropic tested agents on 405 contracts exploited between 2020 and 2025. The agents generated turnkey exploits for 207, representing $550.1 million in simulated stolen funds. Anthropic emphasized that these were simulations, not actual losses. In another test, agents examined 2,849 recent contracts with no known issues and found two previously unknown vulnerabilities, with $3,694 in simulated exploit value.
That small discovery set shows possibility, not how often AI finds flaws in live systems. Anthropic said GPT-5 API use cost $3,476. It also reported that an independent real attacker exploited a fee-withdrawal flaw four days after an agent found it, draining about $1,000 in fees. The report doesn’t say the attacker used AI.

Benchmark results describe simulated exploits, not confirmed losses.
Background and context for AI smart contract security
Traditional scanners look for known code patterns. Agents can potentially test sequences of actions, which matters because some flaws emerge only when contracts, oracles and other protocol components interact. Source verification can provide information about a contract’s code, but the evidence here does not establish that verification alone proves a contract is safe.
Other benchmarks show progress and limits. OpenAI and Paradigm introduced EVMbench in February 2026, using 117 vulnerabilities from 40 audits. OpenAI reported GPT-5.3-Codex achieved 71% success on sandboxed exploit tasks, compared with 33.3% for GPT-5. These results apply to the benchmark’s test conditions.
An October 2026 CyberChainBench study tested agents against 541 exploit incidents across nine EVM-compatible chains. Its best configuration scored 37.5% for detection, 43.7% for exploitation and 23.4% for patching. It reported $57.4 million in simulated exploit profit across 200 post-cutoff cases. OpenZeppelin later judged some high-severity issues not exploitable, and the researchers noted that public training data may exaggerate results.
Instructions can also affect outcomes. In a16z’s test of 20 historical DeFi price-manipulation cases, an isolated agent succeeded in two cases, or 10%. With structured, domain-specific guidance, it succeeded in 14, or 70%. Capability may depend on an agent’s environment and the information it receives.
These tests don’t show that AI is responsible for major on-chain thefts. They do suggest automated tools could help security teams find flaws, while also lowering the effort needed to probe targets. U.Today said smaller protocols may see an uptick in exploit attempts as agents become easier to use.

Automated security testing can probe interactions across protocol components.
Market reaction and reported losses
Available market data doesn’t show that prices reacted specifically to U.Today’s report. At 08:13 UTC on November 19, CoinGecko data showed Bitcoin down 1.31% over 24 hours and Ether down 0.24%. This is market context, not proof of a response to the security story.
Asset
Price, USD
24-hour change
Bitcoin (BTC)
$63,648
-1.31%
Ether (ETH)
$1,801.65
-0.24%
Security loss figures also need context. Hacken’s Q3 2026 report recorded $1.1808 billion in losses across 107 incidents from July 1 to September 28. Application smart contracts accounted for 71 incidents and $228.2 million, or 19% of losses. U.Today cited smart contract exploits as 48% of third-quarter losses. The available information doesn’t reconcile the categories, so the figures shouldn’t be treated as equivalent.
Chainalysis reported about $36.7 million stolen across four hacks of unverified smart contracts in the six months covered by its June 2026 analysis. It identified AI-assisted exploit development as an emerging risk, but didn’t attribute those thefts to AI. On November 19, the Crypto Fear & Greed Index was 10, classified as Extreme Fear by Alternative.me. That indicates broad sentiment, not concern about AI security specifically.
What to watch next
The key question is whether agents can reliably identify fresh vulnerabilities and whether teams can patch them before attackers act. Future benchmarks would be more informative if they disclosed what data agents saw, how much human guidance they received and whether independent reviewers confirmed each flaw was exploitable.
Protocol teams may use agents for defensive testing, but findings need review. Detection isn’t the same as a safe fix, and patches can introduce new bugs. Monitoring, incident response and transaction testing also matter.
The evidence discussed here does not establish that AI caused the reported thefts or determine whether any particular protocol is safe. More meaningful evidence would include independently verified incidents and measured improvement in detection and patching.

Teams still need review, monitoring and a tested response plan.
FAQ
Can AI agents hack smart contracts?
Agents have generated exploits in benchmark simulations and found some previously unknown flaws. This demonstrates capability under test conditions, not widespread AI-led theft from live protocols.
Has AI caused the reported DeFi hacks?
Hacken’s and Chainalysis’s figures don’t establish AI involvement. Chainalysis described AI-assisted exploit development as an emerging risk but didn’t attribute the unverified-contract thefts it analyzed to AI.
How can AI smart contract security improve?
Teams can test code and contract interactions, review agent findings and monitor deployed systems. AI smart contract security is evolving, and benchmark results should be checked against real, independently verified outcomes.
Related Articles
Sources
This article is for informational purposes only and does not constitute financial or investment advice. Cryptocurrencies are highly volatile; always do your own research before investing.
Read More:AI Smart Contract Security: Can Agents Hack DeFi?">AI Smart Contract Security: Can Agents Hack DeFi?