What happened
AWS Security Blog (Aug 27, 2026) detailed a reference architecture using Strands Agents SDK lifecycle hooks (BeforeInvocationEvent, BeforeToolCallEvent, AfterToolCallEvent) to apply Bedrock's ApplyGuardrail API at tool-call and MCP-interaction boundaries, not just at model input/output — closing gaps where model-level guardrails don't reach.
Why it matters
Tool/MCP interactions are exactly the surface Microsoft's concurrent threat research (LiteLLM/RAGFlow/Kestra compromises) showed attackers exploiting; this gives Bedrock-based agent builders a concrete, code-level pattern to enforce guardrails at the tool trust boundary rather than relying solely on model-level filtering.
Applicability
Teams building production agents on Bedrock/Strands should adopt these checkpoints now, especially where agents call external tools or MCP servers with sensitive data access.