or click to navigate
RAXE · STRATEGIC PROSPECTUS

Your AI agents are acting. Can you answer when the board asks how?

Runtime visibility, control, and evidence for your AI agents, inside your boundary, without re-platforming.

GatewayNetwork · provider routing
SDKIn-app · agent & tool context
HostWorkload · process · egress
CloudAWS · CloudTrail · AgentCore
RolloutObserve → Supervise → Enforce
ACME · Illustrative SampleRAXE · Coverage · 62% of AI estate · enf / blk = Enforce phase, target state
In-boundary
Estate · trackedSample
Apps50+3w
Workloads128
Providers18
Assigned92%
Coverageclimbing
62%
Target 85% · Q4
Latency · gatewithin SLO
3.1ms
fast rules · full verdict ~150ms
Runtimehealthy
Req / s147↑
Inspected 5m42.1k
Budget used78%
Application
SDK · in-app context
Coverage84% · 42 / 50
python node typescript go
finance-botfastapi
prompt · summaryenf2.4ms
tool · query_dbenf1.9ms
support-agentnext.js · mcp
mcp · file_readsup3.8ms
chain · 2 subssup2.7ms
Req/s89
P992.8ms
Errors0.02%
Host
Sensor · workload · process
Coverage51% · clusters A/B
k8s docker systemd ec2 lambda
svc-ingest-v2ns: prod-ai
proc · ollama-serveobs
workload · newobs
svc-dev-copilotec2 · systemd
auth · overscopesup
egress · 1/200obs
Workloads128
Shadow AI7
Unowned3
Gateway
Egress · provider routing
Coverage72% · 18 providers
openai anthropic bedrock azure
approved providersallowlist v4
POST openaienf4.1ms
POST anthropicenf3.7ms
shadow providersauto-discover
unknown-llm-proxy.ioblk1.2ms
Egress 5m1.84GB
Would-block12
Budget78%
Rollout
Observebaseline
Superviseapprovals
Enforceonce proven
Observe Supervise Enforce · target state
Illustrative sample · ACME
The Gap

Controls stop at the prompt.
Agents keep going.

Agents now read files, call APIs, chain tools, and move data across sessions. Your risk is not the prompt. It’s everything the model does next with enterprise permissions.

01

Tool-calling turns models into actors

MCP, function calls, and agent frameworks execute against real systems, with your access tokens.

02

Prompt scanners can’t see the action path

Ingress filters rarely explain what happened after the model picked a tool, arguments, and destination.

03

Shadow AI is measurable, and expensive

IBM 2025: shadow AI adds +$670K to breach cost. 97% of AI incidents had no access control. IBM, 2025 →

REQ-ACME-4A8F · Session 0172 · 11:41:03
Illustrative sample · ACME Agent: finance-bot
Prompt
“summarise contract_v4.pdf”
Tool Call
read_file(contract_v4)
Data
PII · customers · 312 KB
Network
→ unknown-llm-proxy.io
Exposure
LEAK
Prompt firewalls end here.

But the material risk: tool, data, destination.

RAXE covers the full path
What Changes

RAXE turns invisible AI activity
into a governed operating model.

Not a prompt firewall. An operating layer that lets senior teams observe, review, govern, and evidence AI activity as adoption scales.

01 / Visibility

Map real AI usage

Identify providers, applications, agents, tools, sessions, and the runtime side effects they actually produce.

02 / Sovereignty
No vendor cloud

Keep scoring in-boundary

Inspect activity inside your environment. No sensitive prompts or tool calls routed to a vendor scanning cloud.

03 / Control
observesuperviseenforce

Move at program speed

Start in observe mode, add targeted review, then enforce where ownership is clear. No re-platforming required.

04 / Evidence
req-4a8freview
req-4a9cwould-block
req-4a1dlog
req-4a0ballow

Create defensible records

Explainable decisions, correlation IDs, SIEM export, and approval history ready for audit and board reviews.

Architecture

Start where you already
have control. Expand over time.

Four surfaces, one shared scoring and evidence model. Begin in observe mode without asking teams to re-platform.

Pilot Entry Point

AI Security Gateway

Provider routes, shadow AI discovery, and centralised reporting at the network layer. Virtual keys, budgets and rate controls: roadmap.

↳ Gateway · base-URL change
Pilot Entry Point

Application SDK

$ pip install raxe
from raxe import trace
trace("tool_call", args)

In-app context: prompts, tools, arguments, permissions, sessions, and inline approvals – supervise phase, target state.

↳ SDK · one import, zero re-platform
Validated – flagship

Host Sensor

Kernel-level eBPF sensor catches in-process file and secret access – invisible to proxies and logs – pinned to the exact agent session.

↳ Host · kernel-observed reality
Early Access

AWS Cloud Activity

CloudTrail · agent-role activity
GetSecretValue prod/payments-db
→ data-exfiltration alert · same queue

Reads your live CloudTrail from inside your environment; agent-role activity lands on the same timeline, correlated to agent sessions – correlation, not exact attribution. Bedrock AgentCore telemetry joins it.

↳ Cloud · read-only, in your account

Most pilots start with Gateway or SDK. The host sensor – validated end-to-end on fresh deploys – adds kernel-observed reality once workload coverage is in scope.

Observe

Prompt · Tool · Session · Process · Network

Capture the signal available at each control point.

Inspect

Fast rules + ML + anomaly signals

Local scoring for known, ambiguous, and novel behaviour.

Verdict

Decision · rationale · evidence

Clear verdict with correlation context and latency.

Action

allow · log · review · escalate · block*

Same operating model across every surface. *Block lands in the Enforce phase, once detections are proven.

REQ-ACME-4A8F · Score Card
Illustrative sample · ACME
Fast RulesHIGH · 0.86
Matches policy.data-boundary: PII leaving approved provider set.
ML ClassifierMED · 0.72
Family: exfiltration · harm-context: confidential.
Out-of-distributionNOVEL · 0.64
Unfamiliar destination + tool pattern for acme-finance-bot session.
VerdictESCALATE
Latencyfast rules 3ms · full verdict 150ms
Three independent signals agree: data-boundary violation, exfiltration pattern, novel destination. Route to approver · attach evidence · continue in supervised posture.
Detection

Transparent rules.
Local ML. Explainable verdicts.

Executives don’t need a black-box alert stream. They need a defensible control model.

  • 01

    Fast rules for known risk

    Transparent patterns for policy triggers, abuse, secrets, and structured-content risks.

  • 02

    Classifier stack for context

    Family, severity, technique, and harm-context: scored locally, never a black box.

  • 03

    Out-of-distribution signals for novelty

    Surface unfamiliar runtime patterns for human review rather than silently passing them through.

  • 04

    Decision logic with rationale

    Every verdict includes contributing signals, latency, correlation IDs, and next action.

Analyst View

Every detection becomes
an investigation record.

Prompt, tool call, session, workload, network route, and final decision connected into one reviewable thread.

Investigation Thread · Illustrative sample · ACME

Confidential customer data → unknown provider

req-acme-4a8f
11:41:03 · Prompt
acme.alex → acme-finance-bot
“Summarise this confidential customer contract”
11:41:04 · Tool call
read_file(contract_v4.pdf)
Permissions: read/approved · Output: 312kb, PII detected
11:41:05 · Data
class: confidential · customer-PII
Destination flagged: unknown-llm-proxy.io
11:41:05 · Verdict
ESCALATE · data-boundary
Rules(0.86) + ML(0.72) + OOD(0.64) → full verdict 150ms · fast rules 3ms at the gateway
11:41:06 · Action
Approver notified · flagged for review – supervise phase, target state
Owner: acme-finance-app · SIEM: splunk-acme-01
11:42:18 · Resolved
Policy update · approved-providers only
Audit record sealed · 73s total review time
Prompt / Response

Who asked, where it went

Provider, model, user, team, redaction events, and prompt-level risk context.

Tool Call / MCP

What the agent tried to do

tool_name, arguments, permissions, output scanning on the real action path.

Session

Whether it escalated

Request & session IDs, escalation trends, cumulative risk across turns.

Process
Validated

Where it ran

Kernel-observed process chain, file access, workload identity, surrounding execution context.

Network
Validated

Where traffic moved

Destination, DNS, byte counts, provider usage patterns for shadow AI governance.

Decision

What was decided

Action, posture, confidence, rationale, latency, correlation for review or export.

Walkthrough · AWS CloudTrail + AgentCore

Your agent’s cloud role reads a production secret.
Who notices?

Which agent role read that secret? CloudTrail knows. Now you do. Agents assume IAM roles, read secrets, and call AWS services under identities your SOC rarely watches – RAXE reads your live CloudTrail and raises detections as agent activity lands.

01 · Agent Role

Unattended agent, real permissions

nightly-reconciliation-agent · IAM role

An overnight agent assumes its cloud role. No human is watching.

02 · Cloud Action

The role reads a production secret

GetSecretValue · prod/payments-db

The read happens inside AWS – no proxy, no SDK, no host in the path.

03 · CloudTrail → RAXE

Detection as activity lands

data-exfiltration alert · agent-role secret read

RAXE reads your live CloudTrail from inside your environment and raises the alert.

04 · Same Console

Reviewed on one timeline

correlated to agent session · one queue

Next to gateway claims and kernel-observed file access. Bedrock AgentCore telemetry joins the same timeline.

API Action
GetSecretValue · prod/payments-db
Source
AWS CloudTrail · your account
Attribution
Correlated to agent session – correlation, not exact attribution
Next Action
Review in console · assign owner

Illustrative sample · self-hosted in your environment, watching your own AWS accounts – no vendor cloud receives your CloudTrail · early access – design partner programme

Operating Model

Observe. Supervise.
Enforce.

A practical adoption path. Begin with telemetry. Add review for higher-risk actions. Enforce once detections are proven and owners are clear.

  • Integrations Splunk, Sentinel, Falcon LogScale, ArcSight – via CEF / syslog / JSON export.
  • Workflow Approval gates and review queues – supervise phase, target state – with audit history embedded in existing ops.
  • Rollout Per-app, per-team, per-policy scope. Toggle posture without re-deploying code.
Adoption Path

One path, three postures.

Choose your starting posture during the 4-week pilot. Toggle without redeploying code.

1
2
3

Observe

Capture runtime telemetry. Export to SOC. Understand real behaviour before changing policy.

Supervised

Approval gates on higher-risk actions. Targeted review without halting everyday usage.

Enforce

Block, redact, or escalate once detections are proven and ownership is clear.

Action vocabulary · block & redact land in the Enforce phase
Allow Log Review Block Escalate
Executive Outcomes

Why leadership teams
buy RAXE.

Not another dashboard. A control model for AI adoption that produces the records leadership, audit, and regulators ask for.

Security Leadership

See your AI risk.

AI-specific telemetry, explainable detections, a path from visibility to control – once detections are proven.

Technology Leadership

Control without rebuilding.

Deploy by base-URL change, SDK install, or host coverage instead of rebuilding the stack.

APP RAXE LLM
AI / Product Leadership

Govern your agents.

Runtime oversight, tool-call review, and policy control without losing deployment speed.

agent.1 agent.2 agent.3 agent.4 agent.5 agent.6 GOVERNED
Risk / Compliance

Prove oversight.

Records of monitoring, review, and control activity that support governance and audit conversations.

REQ-4A8F SEALED REQ-4A9C SEALED REQ-4A1D SEALED
Sovereignty

Your data.
Your boundary.

Scoring runs where your team can govern it. No vendor-operated scanning cloud required for RAXE to inspect AI traffic.

  • Scoring runs inside your VPC, on-prem, or edge. No sensitive traffic to a vendor cloud.
  • Inspect and govern even when the model provider is external.
  • No-egress operation: model bundles locally mirrored and pre-provisioned – no outbound calls required to score.
  • Same scoring contract for embedded and centralised runtime services.
Customer environment
Vendor scanning cloud
Control Points

Gateway · SDK · Host

All sit inside your control boundary. Network, application, and workload coverage share one contract.

Scoring Plane

Embedded or central

Same evidence model whether scoring runs inline with your app or as a central service.

Model Bundles

Locally mirrored

Model bundles are mirrored into your environment; scoring never phones home.

Evidence

Into your SIEM

Decisions and audit history flow to existing pipelines, not to a vendor-owned cloud.

✕No inspection traffic to vendor cloud
Proof of Value

A 4-week pilot
ends with an executive decision.

Walk away with executive findings, evidence exports, and a production rollout plan. One control point, observe mode first.

01
Week 1

Select the control point

Stand up the chosen surface: gateway, SDK, or workload coverage. Begin in observe mode. Wire evidence exports to SIEM.

02
Week 2

Baseline real usage

Review first detections. Correlate runtime evidence. Tune thresholds against observed patterns across teams and apps.

03
Week 3

Test targeted controls

Introduce approval gates or targeted actions where the program owner wants stronger oversight. Measure noise vs. signal.

04
Week 4

Decide next posture

Deliver executive findings, evidence exports, control recommendations, and a production rollout plan.

2–4w
Typical Pilot Window
Observe
Default Starting Posture
SIEM
Evidence Into Existing Workflow
Plan
Executive Rollout Decision

Validation discipline: 2× 12/12 hardened acceptance runs on fresh deploys; audit chain verified Intact – recorded on a live deployment.

Executive Decision
RAXE

Make AI
adoption visible.
Then governable.

RAXE gives senior teams a practical path to identify real AI usage, choose the first control point, and leave the first 30 days with evidence leadership can act on.

01 · Map

Map usage

Teams, agents, providers, and workloads active today.

02 · Control

Choose control

Gateway, application SDK, host coverage, or cross-layer pilot.

03 · Prove

Prove value

Findings, exports, review workflow, and rollout decision.

Next step

Book a 30-minute leadership walkthrough.

We’ll map your first AI control point and define what a 2–4 week proof of value should produce.

01
Confirm the risk lens

Shadow AI, tool actions, data movement, unmanaged workloads.

02
Select the first surface

Gateway, application SDK, host sensor, or cross-layer.

03
Define the evidence pack

Executive summary, exports, workflow, rollout recommendation.

Book a 30-min walkthrough
Pick a time that works for you. No form-fill, no 24-hour wait.
Prefer to leave details?
Leave a work email and optional context. We’ll respond within 24 hours.
Early access – design partner programme No re-platforming Observe first Evidence in 2–4 weeks