Archestra's OpenAPPA Hits 0% Attack Success Rate on Two Security Benchmarks
Archestra released OpenAPPA, an open-source security engine that runs outside the agent's prompt and execution loop to stop data exfiltration from prompt injection or hallucination. It scored a 0% attack success rate on Bench-Corp and AgentThreatBench with 89% task completion, versus 10% attacks for Claude Code auto mode and 31% for Microsoft FIDES.
- 0% attack success rate on Bench-Corp and AgentThreatBench with 89% task completion
- Claude Code auto mode: 10% attacks, 90% tasks; Microsoft FIDES: 31% attacks, 41% tasks
- Policies live in a single appa.toml file: data sources, audiences, trust levels
- Disabling recovery strategies drops task completion to 35%
Read next
Security