Anthropic Publishes Quantified Prompt Injection Failure Rates — First Vendor to Make Agent Security Measurable

VentureBeat·medium signal

Anthropic published prompt injection attack success rates across four distinct agent surfaces — the first AI vendor to provide quantified security metrics for procurement decisions. In constrained coding environments, Claude Opus 4.6 achieved 0% injection success across 200 attempts. However, in GUI-based systems with extended thinking, single-attempt success reaches 17.8%, climbing to 78.6% by the 200th attempt without safeguards (57.1% with them). Promptfoo's independent evaluation found jailbreak rates climbing from 4.3% baseline to 78.5% in multi-turn scenarios.