Agentic Security on AWS: What Nemotron Adds

Security teams don’t lose to “more alerts.” They lose to time.

In 2025, attackers are using AI to compress every step of an intrusion: faster recon, faster phishing iteration, faster lateral movement, faster privilege escalation. Defenders feel that compression as a pileup: more detections, more context to gather, more decisions to justify, and the same number of analysts to do it. The bottleneck isn’t visibility anymore—it’s decision latency.

CrowdStrike’s move to use NVIDIA Nemotron reasoning models through Amazon Bedrock for agentic capabilities inside the Falcon platform is interesting because it targets that bottleneck directly. Not with yet another chatbot, but with the idea that security AI should reason and act, under controls, in the same place the telemetry already lives.

This post is part of our AI in Cybersecurity series, where we keep coming back to one point: AI only helps if it turns messy security data into repeatable actions—and does it safely.

Agentic security: the real promise is fewer “hand-offs”

Agentic security isn’t “AI that writes summaries.” It’s AI that completes a security task end-to-end—from understanding an alert, to gathering evidence, to taking a bounded response step (or teeing up the right next step for a human).

Here’s the simplest way I’ve found to explain it to leaders:

Automation executes. Agentic security decides what to execute next—based on evidence and risk.

That difference matters because most SOC workflows fail in the seams between tools and teams:

• The EDR alert fires.
• Someone pivots to identity logs.
• Someone else checks cloud control-plane events.
• A third person tries to understand “is this normal for this host/user?”
• Then you argue about containment because nobody wants to break production.

Each hand-off adds delay, and delay is exactly what AI-enabled adversaries are trying to buy.

CrowdStrike positions Falcon as an agentic security platform, and in the source announcement, highlights two places Nemotron shows up:

• Falcon Fusion SOAR (orchestration and response playbooks)
• Charlotte AI AgentWorks (building task-specific security agents)

The compelling part isn’t the model name. It’s the focus on reasoning + action in core SOC workflows.

Why Nemotron + Amazon Bedrock is a practical combination

This integration matters for three operational reasons: reasoning quality, deployment speed, and scalability.

Reasoning models reduce the “glue work” in triage

Most SOC effort isn’t spent on exotic reverse engineering. It’s spent on “glue work”—correlating signals that were never designed to line up:

• an endpoint process tree
• a suspicious OAuth consent
• a burst of unusual API calls
• a new persistence mechanism
• a weird DNS pattern

Reasoning-capable LLMs can help connect those signals into a plausible narrative, but narratives aren’t enough. The SOC needs the model to produce structured next actions:

• what evidence to pull next
• what hypotheses to test
• what response steps are safe given the asset’s role

Nemotron is described as an efficient, open model family built to handle text, code, and documents—useful in security because investigations aren’t just log lines. They include scripts, configuration files, tickets, and runbooks.

Bedrock removes infrastructure drag (and that’s not a small deal)

If you’ve ever tried to stand up generative AI in a regulated environment, you already know the hardest part isn’t the prompt. It’s the platform work:

• provisioning inference capacity
• scaling for incident spikes
• controlling data paths
• managing permissions
• handling model updates

Using Amazon Bedrock as the delivery layer (as described in the RSS source) shifts much of that burden to managed services, which tends to speed up adoption inside enterprises—especially those already standardized on AWS.

The result: teams can spend more time on governance and outcomes (what should the agent do?) and less time on undifferentiated plumbing.

Hardware acceleration shows up as “faster decisions,” not “faster tokens”

Security is a latency game. When a human has to wait for an AI system to complete multi-step reasoning, the “agentic” experience collapses.

NVIDIA’s angle is often framed as acceleration, but the SOC translation is straightforward:

• faster reasoning loops
• more concurrent investigations
• lower queue times during major incidents

In practice, that means your automation doesn’t just exist—it’s available when you need it most.

Where agentic AI actually helps: Fusion SOAR and AgentWorks

If you’re evaluating agentic security for lead-worthy reasons (budget, roadmap, measurable ROI), focus on workflows—not features.

Falcon Fusion SOAR: from static playbooks to adaptive playbooks

Traditional SOAR playbooks are brittle. They assume the world is consistent:

• alerts are clean
• entity names match
• enrichment sources respond
• the same steps apply to every asset

The source article claims Nemotron enhances reasoning so playbooks can become more context-aware: understanding alert relationships, prioritizing by risk, and executing complex actions.

That’s the difference between:

• “If alert type = X, then isolate host”

and

• “If alert type = X and the host is a domain controller, don’t isolate; instead block the token, disable the session, and open a high-severity incident with these artifacts attached.”

When adaptive playbooks work, you see two outcomes quickly:

1. MTTR drops because evidence collection and first-response actions happen immediately.
2. Analyst confidence rises because the system shows its reasoning trail and constraints.

Charlotte AI AgentWorks: task-specific agents are more valuable than general assistants

Most companies start with a general security assistant (“Ask it anything”). That’s fine for onboarding and knowledge lookup, but it rarely becomes mission-critical.

Task-specific agents do.

AgentWorks is positioned as a way to create specialized security agents that don’t just answer questions—they act, with “comprehensive awareness” across the environment in real time.

The winning pattern I’ve seen is to define agents around repeatable jobs with clear boundaries, like:

• Phishing triage agent: extract IOCs, check sender infrastructure, correlate with mailbox rules, recommend quarantine scope.
• Identity investigation agent: detect impossible travel + token abuse patterns, validate conditional access logs, recommend session revocation.
• Cloud misconfiguration agent: identify risky IAM changes, map affected resources, open a ticket with least-privilege fixes.
• Ransomware containment agent: detect encryption behavior + lateral movement, propose segmented containment steps based on asset criticality.

The point is control. You can approve agent actions, limit blast radius, and audit decisions.

The safety question: what “agentic” must include to be deployable

Most companies get this wrong: they treat safety as a policy doc.

Agentic security requires safety as a product capability, or you won’t deploy it beyond a pilot. Whether you’re using CrowdStrike’s approach or another platform, insist on these controls.

Minimum controls for agentic SOC operations

1. Human-in-the-loop gates for high-impact actions
   • isolation, account disablement, firewall changes, mass quarantine
2. Least-privilege tool access
   • agents should only be able to call a narrow set of actions, scoped by environment
3. Deterministic guardrails
   • allowlists/denylists, time windows, asset criticality constraints
4. Auditability
   • what data was used, what reasoning was applied, what actions were taken
5. Fallback behavior
   • if enrichment fails or confidence is low, the agent should degrade safely (create a case, request human review)

A useful one-liner to align teams:

If an agent can’t explain what it’s doing, it shouldn’t be doing it.

This matters even more during the holidays (right now) when many orgs run lean staffing, change freezes, and higher fraud/phishing volumes. Agentic automation can help, but only if it’s constrained enough to be trusted at 2 a.m.

What to ask vendors before you buy “agentic security”

If your goal is leads and outcomes—not novelty—use these questions to force clarity.

Evaluation questions that surface real capability

• What actions can the agent take natively, and what requires custom integration?
• How does the system correlate across endpoint, identity, and cloud signals without manual stitching?
• What’s the default approval model for containment actions?
• How are playbooks tested to prevent destructive loops or repeated actions?
• Can we simulate incidents and measure MTTR improvement before rollout?
• How do you prevent data leakage when the model processes sensitive incident context?

Success metrics that actually reflect value

Pick 3–5, baseline them, and track them monthly:

• Mean Time to Acknowledge (MTTA) for high-severity alerts
• Mean Time to Contain (MTTC) for confirmed incidents
• % of incidents with complete evidence packs attached automatically
• Analyst hours saved per week on enrichment and ticket writing
• False positive closure rate (should improve if reasoning is working)

If you can’t measure it, you can’t defend the spend.

People also ask: does agentic AI replace SOC analysts?

No—and if a vendor implies it does, treat that as a red flag.

Agentic AI replaces two things:

• waiting (for enrichment, correlation, and rote steps)
• busywork (copy/paste investigations and repetitive actions)

It doesn’t replace:

• risk decisions (what’s acceptable downtime?)
• adversary tradecraft understanding
• incident command judgment
• stakeholder communication

A strong agentic SOC makes analysts faster and more consistent. It also makes your best analysts easier to scale because their runbooks become executable.

What this signals for the AI in Cybersecurity trend in 2026

The direction is clear: security AI is shifting from “assistant” to “workforce.” Not a marketing workforce—a literal operational workforce with defined jobs, permissions, and audit trails.

CrowdStrike’s use of NVIDIA Nemotron through Amazon Bedrock is a good example of how that shift happens in real products:

• foundation models provide reasoning
• managed platforms provide scale and governance primitives
• security platforms provide the data, context, and controlled actions

If you’re planning your 2026 SOC roadmap, the question isn’t whether you’ll adopt AI-driven threat detection and response. You already are. The question is whether you’ll do it in a way that reduces decision latency without creating a new category of operational risk.

If you’re exploring agentic security and want to pressure-test the approach, start small: pick one painful workflow (phishing triage, identity investigations, ransomware containment), define guardrails, run simulations, and measure MTTR.

What part of your SOC would benefit most from an AI agent that can act, not just explain—triage, investigation, or response?