Straiker finds 94.4% prompt injection

- Agentic security company Straiker published benchmark results claiming 94.4% of tested LLM agents were vulnerable to prompt injection and 83.3% to retrieval-based backdoors. - They also reported 100% susceptibility to inter-agent trust exploits, arguing static guardrails fail at runtime for tool use and objective drift. - The post warns runtime behavior, multi-step workflows, and MCP setups need dynamic defenses, not just pre-deployment filters. (x.com)