Professor releases 548 recorded adversarial calls attacking his AI oral examiner
ipeirotis · x · 2026-11-13
An NYU professor turned his AI product-management final into a voice-agent oral exam using a cloned voice, left it online, and received 1,488 calls in 10 months: 330 silent, about half genuine exam takers, and 548 trying to test, troll or break the examiner.
- The 548 adversarial calls (1,298 attacks) are now a public dataset with full recordings, browsable on the blog and downloadable from Hugging Face and Zenodo
- The post is a highlight reel with timestamped "Listen" links to each quoted moment
- The data stress-tested the examiner agent to reduce fragility in future versions
A rare real-world corpus for anyone studying agent safety and adversarial testing.
More from Safety
- Western University reviews $72 AI tutor after students cry foul over cost, privacy — joshgans · 2026-11-13
- Another Maine town extends data center moratorium as local opposition spreads to 12+ communities — jonippolito · 2026-11-13
- BrowseSafe++ ships major improvements to open-source agentic security benchmark work — marktenenholtz · 2026-11-13
- Anthropic roadshow showcasing Accenture as partner raises doubts about auditor independence — DKokotajlo · 2026-11-13
- Two NVIDIA driver bugs lurked 4+ years, letting GPU tenants root the host — SirJerryLion · 2026-11-13
- Harvard FraudBench: AI fraud rate jumps from 3% to 21.4% as agents act on your behalf — rohanpaul_ai · 2026-11-13