Black Hat USA · Lightning Talk 15 minutes

From Assistant
to Assistant

The evolution of AI at Elastic Security — three years from a chat sidekick to agents that investigate, reason, and act.

James Spiteri · Elastic Security

Quick show of hands

Where is your SOC on this curve?

presenter: click a card per raised hand · shift-click to undo

Three years · three paradigms

Chat → Purpose-built → Agentic

💬

2023 · Ask

ChatGPT moment. Natural language becomes the interface — AI Assistant answers questions, explains alerts, writes queries. Human does everything else.

GPT-4 · 32K tokens
⚙️

2024 · Apply

LLMs aimed at specific painful jobs: Attack Discovery triages alerts in bulk; Automatic Import, Migration & Troubleshooting kill weeks of toil. Output lands in real UIs, not chat.

Claude 3 era · LangGraph
🤖

2025–26 · Act

Agents with tools, reasoning you can read, workflows, MCP & A2A. From answering questions to running investigations end-to-end — with the analyst in control.

Agent Builder · 9.3 → 9.4
The journey · click any milestone

2023 → today, in seven moves

milestone screenshot
June 2023

The chat era begins

ChatGPT had just taken the world by storm — natural language democratized the technology. We shipped the Elastic AI Assistant to find out what value it could actually deliver in a SOC.

Alert summarizationQuery conversion Workflow suggestionsAnonymization built-in
32K
max tokens — GPT-4 was the ceiling
1st
choose-your-own-LLM in a SIEM assistant
Elastic AI Assistant · October 2023 · connector: "James Azure 32K"
Elastic AI Assistant in 2023

The actual 2023 UI — my own connector, 32K tokens and all. Promising, but a long way from carrying real SOC weight.

Meanwhile, the models…

The floor kept rising under us

4K
GPT-3.5
2022
32K
GPT-4
2023 · Assistant launch
200K
Claude 3
2024 · Attack Discovery
tools
Function calling matures
2024–25
reasoning
Extended thinking + MCP
2025
agentic
Frontier models, late 2025
→ dramatic step change

Context windows → tool use → readable reasoning: each capability unlocked the next product era.

Attack Discovery · triage at machine speed
Attack Discovery detail view

Bulk alert triage → discoveries with attack chains, entities, and MITRE ATT&CK mapping.

May 2024 · patent pending

Attack Discovery: beyond chat

LLM-powered triage over alerts in bulk, with findings in a dedicated interface — not a conversation. Claude 3 had just landed, and the results told us we could keep pushing.

🧩 The insight

Chat is an interface, not the product. Put the model inside the workflow and give the output a real home in the UI.

📉 The payoff

Alert fatigue is a volume problem. Machines read 100 alerts the way analysts read one — connecting them into attack chains automatically.

Lesson № 1 · learned the hard way

No evals, no trust

Attack Discovery's results made one thing obvious: quality drifts. New models, new prompts, new data shapes — without evaluations, you're shipping vibes.

LangSmith eval harness Curated attack-scenario datasets Regression-gated prompt changes Model swap in hours, not weeks
prompt × accuracy — same task, wildly different outcomes
Evaluation heatmap of prompt accuracy

Real eval heatmap: adjacent prompt variants swinging from ~95% to ~25% accuracy.

2024 → 2025 · purpose-built AI

The "Automatic" family — LLMs aimed at toil

📥

Automatic Import

Point it at any log source; it builds the custom integration and pipeline. Weeks of onboarding → minutes.

🔁

Automatic Migration

Translates rules & dashboards from your legacy SIEM into Elastic equivalents — the #1 migration blocker, automated.

🩺

Automatic Troubleshooting

Diagnoses Elastic Defend deployment issues automatically instead of hours of doc-spelunking.

🕸️ Under the hood: agents before they were cool

Assistant re-architected on LangChain LangGraph — first agent graph with knowledge sources, sub-agents for ES|QL query generation. Attack Discovery pivoted to agents too.

Automatic Migration · translated rules
Automatic Migration processed rules
Not just us saying it

The industry noticed

🏆

AWS Partner Awards 2024

Global Generative AI Infrastructure & Data Partner of the Year — announced at re:Invent 2024. Among the first 15 partners awarded the AWS GenAI Competency.

re:Invent 2024
🕸️

LangChain: Top LangGraph Agents

Named #2 of the "Top 5 LangGraph Agents in Production 2024" — recognized as one of the first companies anywhere to ship a production AI agent.

LangGraph · LangSmith
🤝

AWS strategic collaboration

Five-year strategic collaboration agreement with AWS (2025) — accelerating agentic AI, MCP, and agent-to-agent interoperability. 2025 GenAI Partner award global finalist.

Bedrock · MCP · A2A
2025 · MCP lands, and the conversation flips

What 500+ customer conversations taught us

“We don't really know how we would use generative AI in our SOC…”2023–24 · almost every conversation
“We want this end-to-end agentic flow to improve our process — how do we do it with Elastic?”2025 · the new default question
0
feedback sessions & interactions in ~18 months
3
asks heard everywhere: control · customization · insight
2
trust prerequisites: secure actions & readable reasoning
Elastic Workflows · YAML-defined, agent-callable
Elastic Workflows editor

Elastic + Keep Keep joined forces with Elastic — automation for agents, agents for automation.

The 10-month rethink

Agent Builder: agents as a platform primitive

Rebuilt with our search & platform teams: agents defined once in the stack — not bolted onto one app.

Custom agents & prompts Tools: ES|QL, workflows, MCP A2A + full APIs Model-agnostic Scoped security boundaries

⚡ Workflows × Agents

Deterministic YAML workflows become tools agents can call — LLM judgment where you want it, rules-based certainty where you need it. GA in Serverless & 9.3.

January 2026 · Elastic 9.3

Agentic, everywhere it counts

🔍 Agentic investigations

Agents that pull context, run queries, and build the picture across your data — reasoning shown, analyst in charge.

🧠 AI summaries in the flow

Entity risk explained in plain language — why this host, why this score, what to do next — right in Entity Analytics.

🛠️ Your agents, your rules

Agent Builder + Workflows + MCP/A2A across the platform: compose the SOC automation you actually want.

Elastic 9.3 · AI entity summary & risk
Elastic 9.3 AI entity summary

9.3: AI-generated entity summaries with recommended actions, inside the analyst workflow.

Elastic 9.4 · Attack Discovery agentic triaging workflow → Slack
Attack Discovery agentic triaging workflow posting findings to Slack

The full loop: Attack Discovery finds it → workflow opens a case → agent investigates → findings + reasoning land in Slack.

Today · Elastic 9.4

The agentic SOC, shipping

⚡ Workflows GA

Native automation in Elastic Security — triage, enrichment, response, case management. No standalone SOAR required.

🧰 Skills for the Elastic AI Agent

Five SOC skills: alert triage, detection rule authoring, entity investigation, threat hunting, anomaly analysis — chained in one investigation. Agent Builder GA in Security.

🧬 Entity Analytics, rebuilt for agents

Precision Entity Identification, Entity Resolution, Dynamic Watchlists, hunting leads — one authoritative record per person, so agents reason over accurate context.

9.5 is around the corner — more skills incoming: detection emulation, binary analysis, alert dedup.

If you remember five things

What it takes to go from assistant to trusted agent

01

Evals before features

Identical-looking prompts swing 95% → 25%. Measure or you're shipping vibes.

02

Chat is an interface, not a product

Aim models at specific, hated jobs and give outputs a real home in the UI.

03

Secure actions or no actions

Scoped tools, workflow guardrails, explicit boundaries. Agents earn write access.

04

Reasoning must be readable

Analysts trust what they can audit. Show the thought process, always.

05

Users want control, not magic

500+ conversations, one signal: customization and insight beat autonomy theater.

James Spiteri · Elastic Security — thank you. Questions: find me after 🤝

15:00
← → navigate · N notes · T timer · O overview · F fullscreen