Updated 1 minute ago
Related News
Nov 11, 2026
AgentCore memory poisoning: what AWS builds in, what its new write-validation pattern adds, and where the demo evidence stops
An AWS Security Blog post from 9 November 2026 separates the memory-poisoning controls AgentCore already offers (IAM namespace conditions, namespace-variable validation, Cedar policy, custom consolidation prompts, the ApplyGuardrail API) from a custom pre-write Lambda interceptor that scores each write and returns ALLOW, QUARANTINE or DENY. The custom layer is a sample you deploy, not a default setting. Its published 11-scenario demo used a keyword-overlap stand-in for embeddings, so the results do not show production accuracy or false-positive rates. Writes that bypass AgentCore Gateway, such as direct SDK calls, are not covered by the interceptor.
AWSAgentCoreAI security
Nov 11, 2026
Bedrock's AgenticRetrieveStream now takes maxAgentIteration 0 or 1: passages only, or one retrieval plus one answer
On November 9, 2026, AWS said Amazon Bedrock Managed Knowledge Base can run retrieval-only and single-shot RAG requests through AgenticRetrieveStream. Setting maxAgentIteration to 0 with generateResponse false returns passages with no foundation model call. Setting it to 1 retrieves once and generates one answer with no planning rounds. Values 2 to 5 keep the existing agentic behavior. The application chooses the mode; the parameter does not route requests automatically. AWS gives no new latency, cost or accuracy measurements, and the passages-only stream differs in shape from a direct Retrieve call.
AWSAmazon BedrockRAG
Nov 10, 2026
Google's MKA benchmark tests AI agents on 200 private professional tasks; a separate set of 10 public tasks is available for inspection
Google DeepMind's new Multimodal Knowledge Agent benchmark evaluates AI agents on 200 private tasks requiring visual analysis of complex professional files like architectural blueprints and engineering schematics. While the benchmark's headline scores are based on these private evaluations, only 10 sample tasks are publicly available for inspection. The benchmark uses weighted criteria where harder 'Climb' criteria measure advanced performance, and costs reported by Artificial Analysis exclude judge calls, limiting direct workplace cost comparisons.
AI benchmarksAI agentsGoogle DeepMind