What it is
100 hand-authored multi-turn security-audit transcripts with per-row randomized tool names. Built to test whether models learn tool binding and vulnerability-audit behavior, not fixed APIs. 100 multi-turn audit transcripts over 100 unique fake-but-realistic codebases, with 4-5 declared tools per row, randomized tool names, tool calls, shell/file actions, and proposed fixes.
Key results
- 100 transcripts across 100 unique codebases
- 941 tool calls; median 8 per transcript
- MIT license; 127 distinct tool names
Useful for training and evaluating agentic vulnerability-audit models that must read and use per-prompt tool APIs.