Deep Dive · Agentic Triage

An AI analyst that shows its work.

Every alert is triaged by an agent running Cisco's Foundation-sec model on Nvidia hardware, on the endpoint itself. Work that used to sit in a tier-two analyst's queue for an hour is done in seconds. You get the verdict, the pivots it looked at, the evidence it cited, and the ATT&CK technique it mapped. If you disagree, you override it, and the correction trains the next verdict.

The Model

Purpose-built for security, not repurposed from a chat product.

Most AI features bolted into SIEM products rely on general-purpose LLMs. Clever, but not tuned for adversary reasoning. Edge runs Cisco Foundation-sec, a security-tuned model, on Nvidia-class hardware on the endpoint. No data leaves the customer boundary. The model can be audited, swapped, or fine-tuned on your own corpus without sending a single event to a vendor.

Cisco Foundation-sec

Security-tuned reasoning model that understands ATT&CK techniques, attacker tradecraft, and forensic chains of evidence as first-class concepts, not as generic text.

Nvidia on-device inference

Runs on the endpoint itself. No cloud round-trip. No tokens billed to your account. No data egress for any sovereign, regulated or air-gapped estate.

Framework-agnostic fallback

CPU inference with a quantised Llama-3-8B fallback for constrained estates or edge sites where GPU hardware isn't viable. Same verdict format, same evidence chain.

Fine-tunable on your corpus

The model can be adapted to your environment, your normal, your rule library. Your analyst overrides become the training data. The longer you run it, the sharper it gets.

What Your Analysts See

Every verdict arrives with its receipts.

An AI answer is only useful if your analyst can replicate the reasoning in thirty seconds. Every Edge verdict ships with a cited evidence chain, an ATT&CK technique, a confidence score, and the exact telemetry events it pivoted through. There is no black box.

verdictCONFIRMED MALICIOUS · 95% confidence
techniqueT1055.012 · Process Hollowing (Execution)
parentexplorer.exe (legitimate, signed)
childunsigned region within hollowed memory space
outboundC2 IP matched actor TAG-RUBY
blast radius4 endpoints, same tenant, 5-minute window
recommendedhost quarantine · credential revocation · IOC sweep
autonomous?no, analyst approval required

The verdict card is not a PDF export. It's the live working memory of the agent. Click any pivot and you're taken to the event that drove that conclusion. Click the technique and you see every rule in your library mapped to it. Click the actor and you see every other alert across the tenant touching the same IOC.

The Learning Loop

Analyst overrides train the next verdict.

The model does not stay static. Every override your analysts produce, whether confirming, downgrading, or reclassifying a verdict, becomes labelled training data for the next cycle. The platform compounds in sharpness the longer it runs in your environment.

01

Alert arrives

Detection engine fires on a rule or anomaly. The agent pulls cross-source context (endpoint, network, cloud, identity) from the local lakehouse.

02

Agent reasons

Foundation-sec proposes a verdict with technique mapping, evidence chain, and confidence score. Anything below configured threshold routes directly to human.

03

Analyst reviews

Cited evidence is one-click reproducible. Analyst confirms, overrides, or reclassifies. Disagreement is logged with reason code.

04

Correction feeds training

Override signal becomes labelled data for the next model update. Tenant-scoped fine-tuning keeps every customer's learned patterns inside their own boundary.

05

Next verdict arrives sharper

The noise class that yesterday sent a tier-two analyst into a rabbit hole is today closed by the agent with a one-line summary. Humans keep the interesting work.

Scope

Where the AI triages, and where it holds back.

AI is autonomous for

  • Alert enrichment and event stitching
  • Technique mapping to MITRE ATT&CK
  • Evidence chain construction
  • Confidence scoring and false-positive triage
  • Hypothesis suggestion for hunt queries
  • Cross-tenant IOC correlation (respecting isolation)

AI holds back on

  • Response execution beyond ransomware containment
  • Credential revocation and account disablement
  • Firewall or DNS blocks
  • Endpoint reimage
  • Anything with a configured confidence floor below your threshold
  • Anything touching a tenant under investigation hold

Privacy is architectural

Because inference runs on the endpoint and customer telemetry never leaves the customer's lakehouse, there is no vendor data-sharing agreement to renegotiate, no cross-customer contamination risk, and no regulatory surprise when your DPO asks where the model saw your alerts. The answer is: nowhere.

Economics

Hour to seconds. Tier-two to tier-one capacity.

The measurable business impact of agentic triage is analyst hours reclaimed. A conservative baseline: tier-two enrichment and triage work that previously held an analyst for 45–90 minutes per alert now closes in under 30 seconds of wall-clock time, with the analyst reviewing rather than assembling.

< 30s
Alert-to-verdict
95%
Typical confidence on cited verdicts
100%
On-device · no vendor data egress

For MSSPs, this reshapes the book. The margin compression that per-GB cloud SIEM puts on analyst-hours is replaced by agent-assisted tier-one coverage that scales horizontally with tenants, not linearly with headcount.