Lost in Translation How Multilingual Gaps Expose Agentic AI to Real World Risks скачать в хорошем качестве

Lost in Translation How Multilingual Gaps Expose Agentic AI to Real World Risks

🌍🔐 Lost in Translation: Multilingual Gaps in Agentic AI Security This session from the OWASP GenAI Security Project Virtual Summit (October 2025) reveals a hidden attack surface in agentic AI: non-English inputs. Omar Hoffman (Fujitsu Research of Europe) shows how the same prompt injection blocked in English can succeed in other languages (e.g., Hebrew/Hindi), triggering unsafe actions, privacy leaks, and real financial risk—because “English safety” ≠ multilingual safety. Introducing MAPS, a multilingual benchmark suite for agentic AI across reasoning, security, math, and code (≈10,000 tasks / 11 languages). Results show performance drops outside English and attack success rates rise, driven largely by planner drift in step one—where the agent’s plan diverges and everything downstream fails. Key takeaways for red teams: test strategically chosen languages + styles, monitor planner/tool calls, don’t rely on translation as a full fix, and add agent-level guardrails. 👉 Learn more about the OWASP GenAI Security Project: https://genai.owasp.org #AgenticAI #AISecurity #OWASP #GenAI #RedTeaming #PromptInjection #MultilingualAI