TechPrufer
Get started
Elastic
Elastic logoElastic · Feature trail

Agentic AI Context Layer

21 posts · Jan 2024 – Sep 2026

What this trail covers

Elastic's platform supports securing AI systems and environments by providing visibility into AI-specific risks, such as prompt injection, model abuse, untrusted model onboarding, agent manipulation, RAG data poisoning, and safety control tampering. This includes Elastic AI Agent, Workflows, LLM security detections, and agentic SOC platform capabilities. The Elastic 9.4 release enhances Workflows with native automation for triage, enrichment, response, and case management, and introduces Agent B. This post emphasizes the need for visibility and context to detect unusual behavior and respond at machine speed in the context of AI agents in banking cybersecurity, highlighting the importance of correlating signals across applications, APIs, identities, endpoints, networks, and cloud environments to distinguish normal automated activity from threats. It also discusses the evolution of the SOC to combine human judgment with automated investigation and response, and the need for governance and transparency in autonomous systems.

All Elastic trails
  1. AI agents and banking cybersecurity: Why visibility mattersThis post details the cybersecurity challenges posed by AI agents in banking, emphasizing the critical need for visibility and context to detect and respond to threats at machine speed. It highlights how agentic AI introduces autonomy and scale, requiring defenders to correlate signals across diverse data sources (logs, metrics, traces, security events, threat intelligence, identity) to identify patterns. The post discusses the evolution of the SOC to incorporate agentic security for investigation and response, and stresses the importance of governance and transparency for autonomous systems, advocating for a unified data platform to bring together security, observability, and search capabilities.
  2. Reducing LLM hallucinations with atomic claim verificationThis post introduces a novel approach to mitigating LLM hallucinations in knowledge base consolidation by implementing a two-pass system. The first pass uses an LLM to merge articles, with strict constraints to preserve technical payloads verbatim. The second pass, also using an LLM but in a separate context, decomposes source articles into atomic claims and verifies their presence and fidelity in the merged output. This atomic verification process converts open-ended generative judgments into narrow entailment checks, ensuring that omissions and discrepancies are explicitly identified and addressed by human reviewers, thereby improving the trustworthiness and accuracy of AI-generated content.
  3. SNAP payment error detection: how Elastic helps US states beat the FY2028 penaltyThis post details how Elastic's platform can be used to detect and prevent fraud in government programs like SNAP by leveraging layered detection mechanisms including customizable rules, machine learning for behavioral drift, and conversational investigation via Elastic Agent Builder. It emphasizes real-time detection over batch processing to address compliance penalties and improve program integrity.
  4. Why your AI bill tripled while token prices fell 75%This post identifies that agentic AI costs are primarily a control problem due to extended reasoning loops, multi-turn dialogues, and multi-hop data retrievals, leading to increased token consumption. It highlights the lack of per-task cost telemetry and observability as a key issue, contrasting with the falling token prices. The post proposes capturing five fields per agent run (tokens in, tokens out, model used, retry count, task completion) to compute a real cost per completed task and suggests routing traffic through a single gateway for unified cost data. It advocates for measuring the task, not the token, by focusing on metrics like cost per completed task, context efficiency, quality-adjusted efficiency, and business value multiple.
  5. Defending against AI-fueled social engineeringThis post details how AI is accelerating social engineering attacks (spear phishing, smishing, deepfakes) by lowering cost and increasing personalization. It argues that legacy detection methods are insufficient and proposes a defense strategy based on unified telemetry for behavioral analytics and agentic investigation/response. The post emphasizes the need for transparent, model-agnostic AI in defense mechanisms.
  7. How SLED teams win with agentic AI cybersecurityThis post details how SLED organizations can leverage agentic AI for cybersecurity. It highlights the challenges of fragmented security data and how an agentic SOC can accelerate investigations by correlating alerts, gathering context, and preparing response plans. The post emphasizes Elastic's human-in-the-lead model for agentic SOCs, where agents present plans for analyst approval before action. It provides case studies from California's EDD and Arizona's Department of Homeland Security, as well as Vrije Universiteit Brussel and Texas A&M University System, to illustrate how agentic AI extends security team capacity and accelerates threat investigations. The importance of a data mesh architecture for holistic visibility without centralizing sensitive data is also discussed, aligning with Five Eyes guidance on operational visibility for AI agents.
  8. 7 lessons for IT leaders on using observability to monitor AI applicationsThis post contributes lessons learned on monitoring AI applications, specifically focusing on the need for LLM observability. Key technical takeaways include: 1. Implementing LLM observability from the MVP stage to capture usage history for ROI calculation. 2. Utilizing probabilistic metrics like token consumption, retrieval quality, and use-case classification alongside traditional metrics for AI systems. 3. Adopting OpenTelemetry semantic conventions for generative AI to standardize telemetry vocabulary and avoid rebuilds when frameworks or models change. 4. Ensuring instrumentation reports actual provider-billed figures (e.g., token counts) rather than estimations to avoid cost discrepancies. 5. Managing observability costs by avoiding high-cardinality tags (like conversation IDs) and uniform sampling, and instead computing derived values at query time. 6. Debugging AI quality issues by capturing retrieval data (what was retrieved, ranking, support for the answer) to differentiate between model, ranking, and content problems. 7. Emphasizing the creation of robust evaluation systems for AI quality assurance.
  9. From retrieval to agents: 5 takeaways on production architecture for AI agentsThis post elaborates on the evolution of AI agents from retrieval-based systems to more complex agentic architectures. It introduces 'context engineering' as the critical discipline for ensuring agent accuracy, contrasting it with prompt engineering and highlighting the problem of 'context rot.' The post emphasizes the need for real-time data ingestion, hybrid retrieval, reranking, and secure access controls for production agentic workloads. It also discusses the cost implications of large context windows versus efficient retrieval and the drawbacks of stitching together point solutions, advocating for a unified platform approach. Key takeaways include the shift in evaluation criteria for AI platforms, the defining challenge of accuracy, the foundational role of search, and the hidden costs of fragmented AI architectures.
  10. Why agentic treasury needs search and observabilityThis post extends the 'Agentic AI Context Layer' by detailing how search and observability are critical for agentic treasury applications. It emphasizes the need for a unified context layer that combines structured and unstructured data for AI agents, leveraging Elasticsearch for hybrid retrieval. It also highlights the role of observability in validating AI-driven financial decisions by providing operational telemetry and tracing the provenance of recommendations. The post introduces the concept of 'agentic treasury' and the governance implications for AI agents operating in financial contexts.
  11. Building trusted agentic AI in financial services: From data to autonomous actionThis post discusses the shift from generative AI to agentic AI in financial services, emphasizing the critical need for trusted context, governance, and observability. It details how observability (metrics, traces, logs) becomes the control plane for governing AI, enabling reconstruction of reasoning and actions. The post also highlights the role of enterprise search in providing current, permission-aware, and governed knowledge to AI agents, and the evolution of security operations to an agentic model for investigating and correlating AI-driven threats. It frames these capabilities as essential for building trusted agentic AI in regulated industries.
  12. Closing the AI gap in government: Next-gen knowledge access | ElasticThis post elaborates on the challenges and architectural requirements for scaling AI in the public sector, specifically focusing on the need for a governed retrieval layer for AI agents. It introduces the concept of 'next-generation federated knowledge access' which leverages hybrid search (lexical and semantic) and open integration protocols to connect AI agents with authoritative knowledge while maintaining sovereignty, auditability, and public trust. The post emphasizes that operationalizing AI is fundamentally a data readiness problem, contrasting legacy data warehouses with the requirements of agentic AI. It highlights the importance of architectural decisions around data residency, safety guardrails, auditability, and human-in-the-loop approvals. The post also discusses the mission outcomes agencies can expect, such as a unified operational picture, efficiency gains, reduced compliance exposure, faster time to production for AI use cases, and strategic autonomy over AI roadmaps.
  13. Elastic and OpenAI collaborate to bring frontier intelligence to unstructured enterprise dataThis post details the expanded collaboration between Elastic and OpenAI to bring frontier intelligence to unstructured enterprise data. It focuses on how Elasticsearch serves as a context layer for agentic AI, enabling context-aware AI agents that retrieve accurate, permission-aware enterprise knowledge at scale with fewer tokens. It highlights the practical challenge of retrieval, mentioning Elasticsearch's combination of lexical and vector search, semantic reranking, filters, and access controls, citing internal retrieval tests that delivered 0.89 recall and complete data protection. The post also discusses economic considerations, noting that precomputed Knowledge Indicators reduced input-token use by up to 75% in an experiment while increasing answer accuracy. It introduces Elastic Agent Builder's role in improving context and delivering relevance and efficiency with OpenAI models. For observability, it explains how Elastic Observability brings AI application signals together with traditional telemetry, enabling agentic investigations and remediation. For security, it describes Elastic Security as an agentic security operations platform that unifies SIEM, XDR, and automation, using OpenAI's models to correlate alerts into attack chains and assist with triage and response. It mentions Visa's experience cutting triage time and Airtel's faster triage and analysis. The post also touches on Elastic's integration with OpenAI's Compliance Logs Platform and future plans for GPT-5.5 Cyber specific models and general availability of Knowledge Indicators.
  14. Rethinking the SOC: From tool procurement to platform architectureThis post argues for a shift from a tool-procurement model to a platform architecture for Security Operations Centers (SOCs) to address the speed and fidelity of AI-powered threats. It highlights the 'fragmentation tax' of disconnected tools, the limitations of layering AI on top of fragmented data, and the architectural requirements for an AI-era security platform, including unified telemetry, open standards, query-in-place capabilities, model-agnostic AI with transparent reasoning, native automation, and deployment flexibility. It introduces the concept of agentic operations where AI investigates and stages responses, with human analysts reviewing and approving actions, thereby compressing incident timelines and shifting metrics from alerts triaged to attacks neutralized.
  15. Elastic achieves the AWS AI Security Distinction, securing AI-specific risksThis post details Elastic's achievement of the AWS AI Security Distinction, highlighting their five years of security innovation. It emphasizes the integration of Elastic Security with AWS services (GuardDuty, Security Hub, Security Lake, Inspector, Bedrock) and introduces new capabilities in Elastic 9.4, including generally available Elastic Workflows (combining scripted automation and agentic reasoning), Agent Builder, and four new Entity Analytics capabilities (Precision Entity Identification, Entity Resolution, Dynamic Watchlists, and Entity-Driven Hunting Leads). These advancements aim to secure both AWS infrastructure and AI applications by providing visibility into AI-specific risks and enabling faster, more secure AI innovation.
  16. The future of governing AI agentsThis post introduces a four-layer architectural separation for governing AI agents: Skills, Reasoning, Models, and Context. It proposes extracting reasoning into an explicit layer independent of the LLM to ensure predictability, consistency, and explainability, which are crucial for progressive trust and regulatory compliance. The post details how this separation allows trust evidence to attach to the reasoning methodology rather than the model, enabling easier model swapping and continuous evaluation based on classification accuracy, planning quality, retrieval quality, and grounding quality.
  17. Elastic and Cursor partner to accelerate context engineering with coding agentsThis post introduces the integration of Elastic's platform capabilities as first-class Agent Skills and an Elastic Docs MCP server within the Cursor AI coding platform. This enables coding agents to leverage live operational data (observability, security, search) for context engineering, code suggestions, refactoring, and fixes. The integration allows agents to query logs, triage alerts, and access Elasticsearch syntax directly from the editor, enhancing the agentic development workflow with real-time production data. Key contributions include the packaging of Elastic's platform expertise as composable and modular Agent Skills (covering Elasticsearch APIs, Kibana, Elastic Observability, Elastic Security, and Elastic Cloud) built on the agentskills.io specification, and the Elastic Docs MCP server for interacting with Elastic documentation. This integration aims to accelerate product innovation, reduce operational toil, and enable more robust AI solutions at scale by grounding agent suggestions in live operational data.
  18. The big ideas behind retrieval augmented generationThis post explains the core concepts behind Retrieval Augmented Generation (RAG), including the challenges of Large Language Models (LLMs) like hallucination and outdated data. It details how semantic search, powered by vector embeddings, addresses these limitations by enabling LLMs to access and utilize private data. The post contrasts RAG with training LLMs from scratch and fine-tuning, highlighting RAG's practicality and scalability. It also touches upon the cost implications of token usage in LLM interactions and the importance of efficient data retrieval for prompt engineering.
  19. GenAI for customer support: Explore the Elastic Support AssistantThis post announces the general availability of the Elastic Support Assistant, a generative AI powered chat experience for customer support. It details how the assistant uses generative AI and a retrieval augmented architecture to answer product questions by summarizing content from Elastic's blogs, product documentation, and Knowledge Base. The post outlines access methods, use cases (troubleshooting configurations, performance tuning, upgrades, security/compliance, monitoring/alerting, custom use cases/integrations), and provides examples of internal usage, including synthesizing answers across documents and step-by-step guidance.
  20. Reducing false positives with automated SIEM investigations from Elastic and TinesThis post details how Elastic's InfoSec team uses Tines to automate the initial investigation of SIEM alerts, specifically focusing on reducing false positives. The core technical contribution is the implementation of an automated triage workflow that queries Elasticsearch using the _search API based on alert tags (e.g., 'Triage:Asset', 'Triage:PMFA', 'Triage:Workstation'). These queries check for activity from trusted sources (managed workstations, internal network zones, authorized third-party applications, phishing-resistant MFA) to automatically close benign alerts. Alerts that do not meet these criteria are escalated for human investigation. The post highlights the use of the Signals API to set alert status and the Alert Actions feature in Elastic Security to send alerts to Tines via a Tines connector or Webhook. It quantifies the impact, stating that over 3,000 alerts are triaged daily without human interaction, preventing the need for an estimated 94 additional SOC personnel.
  21. Understanding the approximate nearest neighbor (ANN) algorithmThis post provides a detailed explanation of Approximate Nearest Neighbor (ANN) algorithms, which are fundamental to efficient vector search. It covers the definition of ANN, how it works through dimensionality reduction and indexing, and when it's most applicable (large datasets, high-dimensional data, real-time applications). The post also discusses the importance of ANN in vector search for enabling fast retrieval and scalability. It then delves into specific ANN algorithm types: KD-trees, Locality-Sensitive Hashing (LSH), Annoy, and Linear Scan, outlining their characteristics, advantages, and disadvantages. This technical deep dive into ANN algorithms directly supports the underlying search capabilities required for the Agentic AI Context Layer by explaining how to efficiently find similar data points in large, high-dimensional vector spaces.
  22. How Elastic AI Assistant for Security and Amazon Bedrock can empower security analysts for enhanced performanceThis post details the integration of the Elastic AI Assistant for Security with Amazon Bedrock, specifically using Anthropic Claude 2. It highlights how this integration empowers security analysts by enabling natural language interaction for tasks such as alert investigation, incident response, and query generation. The post emphasizes the importance of organizational context for LLMs and how Elastic's platform, combined with Bedrock, provides this context to ensure accurate and relevant AI-generated responses for security use cases.