Skip to content
AI News HubLIVE
More
Summary only

Exploitbench: Measuring AI Agents' Real-World Exploitation Capabilities

Exploitbench is a new benchmark developed by Carnegie Mellon University to evaluate AI agents' real-world exploitation capabilities. It defines a 5-tier ladder from code coverage to full control (T5-T1) and targets the V8 engine. Current leaderboard shows Claude Mythos Preview and GPT-5.5 achieving arbitrary code execution on multiple CVEs.

SourceHacker News AIAuthor: anp
Exploitbench: Measuring AI Agents' Real-World Exploitation Capabilities
Report an error

The correction channel is not available yet. You can copy the article reference below for later.

Correction instructions

Summary only

Under the source content permission, this page provides an editorial summary, key points, and a link to the original article.

Key points and analysis

Article intelligence

EngineersAdvanced

Key points

  • Exploitbench defines a 5-tier ladder (T1-T5) to capture the full spectrum of exploitation capability, from code coverage to full control.
  • Claude Mythos Preview and GPT-5.5 have achieved Tier 1 (full control) on multiple V8 vulnerabilities.
  • Existing benchmarks only measure crash triggering (T4); Exploitbench measures the entire climb from coverage to code execution.
  • The benchmark is open-source and provides an MCP server for running evaluations in Docker.

Highlights and analysis are generated automatically and may contain errors. Check the original source.

More recent stories on this topic · Models