Tuesday, August 11, 2026
LIVEThe Unrelenting Cyber Battle: Hacking Threats and the Imperative of Robust Data Protection///Navigating the Cyber Labyrinth: Bolstering Defenses Against Evolving Hacking Threats///The Dual Front War: Battling Hacking and Bolstering Data Protection in the Digital Age///The Ever-Evolving Cyber Threat Landscape: Navigating Hacking and Fortifying Data Protection///The Unseen Battle: Fortifying Data in an Age of Relentless Hacking///The Unseen War: Hacking's Relentless Advance and the Imperative of Data Protection///The Evolving Threat Landscape: Hacking, Data Protection, and the Imperative for Proactive Security///Navigating the Digital Minefield: Bolstering Data Protection in an Era of Relentless Hacking///The Dual Fronts of Digital Defense: Combating Hacking and Fortifying Data Protection///Hacking's New Frontier: Fortifying Data Protection in the Age of Advanced Cyber Threats///The Dual Front: Navigating Hacking Threats and Fortifying Data Protection in the Digital Age///Navigating the Digital Gauntlet: The Evolving Nexus of Hacking and Data Protection///The Unrelenting Cyber Battle: Hacking Threats and the Imperative of Robust Data Protection///Navigating the Cyber Labyrinth: Bolstering Defenses Against Evolving Hacking Threats///The Dual Front War: Battling Hacking and Bolstering Data Protection in the Digital Age///The Ever-Evolving Cyber Threat Landscape: Navigating Hacking and Fortifying Data Protection///The Unseen Battle: Fortifying Data in an Age of Relentless Hacking///The Unseen War: Hacking's Relentless Advance and the Imperative of Data Protection///The Evolving Threat Landscape: Hacking, Data Protection, and the Imperative for Proactive Security///Navigating the Digital Minefield: Bolstering Data Protection in an Era of Relentless Hacking///The Dual Fronts of Digital Defense: Combating Hacking and Fortifying Data Protection///Hacking's New Frontier: Fortifying Data Protection in the Age of Advanced Cyber Threats///The Dual Front: Navigating Hacking Threats and Fortifying Data Protection in the Digital Age///Navigating the Digital Gauntlet: The Evolving Nexus of Hacking and Data Protection///
Subscribe
Cyber Security
Independent · Digital
Thehackingpost
RSSSubscribe
CybersecurityAI-assisted

OpenAI Launches EVMbench to Detect, Patch, and Exploit Vulnerabilities in Blockchain Environments

OpenAI, in collaboration with Paradigm, has released EVMbench, a benchmark designed to assess AI agents' capabilities in identifying, repairing, and exploiting high-severity vulnerabilities in smart contracts. This benchmark is crucial for evaluating AI…

WBy William HayesFeb 19, 20261 min read

OpenAI, in collaboration with Paradigm, has released EVMbench, a benchmark designed to assess AI agents' capabilities in identifying, repairing, and exploiting high-severity vulnerabilities in smart contracts. This benchmark is crucial for evaluating AI performance in economically significant environments, given that smart contracts secure over $100 billion in crypto assets.

EVMbench leverages 120 curated vulnerabilities from 40 security audits, mostly sourced from open code audit competitions. It also includes scenarios from the Tempo blockchain's security audit process, targeting payment-oriented smart contract code.

EVMbench evaluates AI agents through three modes, each focusing on a different phase of the smart contract security lifecycle:

Detect: Agents audit smart contracts and are scored on the recall of vulnerabilities and audit rewards. Patch: Agents modify vulnerable contracts to preserve functionality, verified through tests and exploit checks. Exploit: Agents perform fund-draining attacks in a sandboxed environment, graded through transaction replay and verification.

EVMbench leverages 120 curated vulnerabilities from 40 security audits, mostly sourced from open code audit competitions.
William Hayes · Thehackingpost

A Rust-based harness supports reproducible evaluations by deploying contracts deterministically and restricting unsafe RPC methods. All exploit tasks are conducted in an isolated environment.

Performance on EVMbench varies across tasks. In exploit mode, GPT‑5.3‑Codex scored 72.2%, significantly higher than GPT‑5's previous score of 31.9%. While agents excel in exploit tasks, detect and patch modes remain challenging due to incomplete audits and difficulty in removing subtle vulnerabilities.

EVMbench does not fully represent real-world smart contract security complexities and cannot distinguish between true vulnerabilities and false positives beyond the human-auditor baseline. OpenAI has allocated $10 million in API credits to its Cybersecurity Grant Program to advance defensive security research.

Advertisement

Additionally, OpenAI announced an expansion of its security research agent, Aardvark, through a private beta program. EVMbench's components are publicly available for further research into AI-driven cybersecurity capabilities.

Based on reporting by Cyber Security News.

AI transparency. This article was produced with the assistance of artificial intelligence and published under human editorial oversight. AI systems can make mistakes. Read how we use AI (EU AI Act, Art. 50).
Related Stories
Cybersecurity

Why Fintech Banking Platforms Are Reducing Operational Costs by 40%

Paige Monroe · Mar 26, 2026
Cybersecurity

Dogecoin Price Prediction: SHIB Burns 172M Tokens as Pepeto Exchange Earns From Every Trade

Eric Wallace · Mar 26, 2026
Cybersecurity

Solana Price Prediction: SOL Holds $86 With Commodity Status as Pepeto Targets Bigger Returns

Zachary Burns · Mar 26, 2026
Cybersecurity

Open Directory Malware Campaign Uses Obfuscated VBS, PNG Loaders and RAT Payloads

Joseph Cain · Mar 26, 2026