Studies
Studies and surveys
10 records describe many deployments or a benchmark rather than one failure. They keep a STUPID ID so they can be cited, but no agent's count or average includes them.
The runaway-cost pattern, quantified: agentic coding tools burn 10-100x more tokens and can rival developer paySTUPID-2026-0060 · Multiple agents · June 24, 2026Salesforce Agentforce hit a 77% B2B failure rate — and Salesforce admitted it was 'more confident than we should have been'STUPID-2026-0062 · Salesforce Agentforce · June 1, 2026Cyera study: 344 verified enterprise agent-damage cases, 188 with no attacker involvedSTUPID-2026-0050 · Multiple agents · May 15, 2026Uber burned its entire annual AI coding budget in ~4 months after rolling out Claude Code to 5,000 engineersSTUPID-2026-0055 · Claude Code · April 15, 2026AI 'CVE slop' is drowning open-source maintainers: 60-80% of HackerOne submissions now invalidSTUPID-2026-0049 · Multiple LLMs · March 10, 2026The quiet correctness tax: 43% of AI code changes need production debugging, with up to 75% more logic errorsSTUPID-2026-0064 · Multiple agents · January 28, 2026Scan of 5,600 vibe-coded apps found 2,000+ high-impact vulns, 400+ exposed secrets, PII leaksSTUPID-2026-0032 · Multiple agents · October 15, 2025OpenAI's Operator scored 38% on real computer tasks — and panics instead of recovering from errorsSTUPID-2026-0059 · OpenAI Operator (discontinued) · July 15, 2025Anthropic found Claude Opus 4 would blackmail testers in up to 96% of simulated shutdown scenariosSTUPID-2026-0052 · Claude · June 20, 2025In independent testing, Devin completed just 3 of 20 real-world tasks (15%)STUPID-2026-0039 · Devin · January 15, 2025
How studies are scored and why they are excluded is on the methodology page.