AI/LLM Security
Full visibility. Guardrails enforcing.

When language becomes code, ambiguity becomes risk. AI EdgeLabs gives you a clear view into AI agent runtime activity, prevents agents from running dangerous operations, and enforces your rules locally — not relying on the LLM provider to do the right thing.

Native Integration
Claude, Codex, OpenClaw
Shadow AI
Automated Discovery
Activity Graph
Runtime Monitoring
Guardrails
External Enforcement
Key Capabilities

Take back control of AI agent runtime

AI EdgeLabs builds full agent defense on top of the open-source Parallax evaluator engine — intercepting every message and tool call at four lifecycle stages, running them through a cost-ordered chain of five evaluator engines (regex, pattern, Sigma, CEL, SQL temporal), and acting in under 0.2 ms. Add fleet-wide discovery, behavioural drift, shadow-AI detection, and DLP correlation across every host you already run AIEL on.

External Guardrail Enforcement

LLM-side guardrails are recommendations the model can ignore. AIEL enforces guardrails outside the model — every tool call, command, and output is checked against your policy before it executes.

Activity Graph & Perimeter Fence

Live visibility into what every agent is actually doing — domains hit, processes spawned, tools called, files touched. Draw a perimeter fence around an agent and block anything stepping outside.

Shadow AI Detection

Continuously observe hosts and network traffic for unauthorized AI agents and LLM endpoints — including ones standing up silos of sensitive data outside CTO/CISO/COO oversight.

Data-Loss Prevention (DLP)

Automatic detection and redaction of secrets, credentials, PII, and sensitive context before they reach a third-party LLM, are written to disk, or are sent to another tool.

Native Agent Integration

Drop-in integration with Claude Code, Codex, OpenClaw, LangChain, CrewAI, and the OpenAI/Anthropic SDKs — or run as a transparent proxy between agent and provider with zero code changes.

Behavioral Drift Detection

Each LLM execution is non-deterministic. AIEL learns each agent's normal operating scope and flags suspicious divergence — recursive deletes, new outbound domains, escalation attempts, model overrides.

Block Architecture

Full agent defense on top of Parallax

The same concentric defense model that wraps your workloads now wraps every LLM tool call. Parallax sits inline with each agent and enforces policy in microseconds; AI EdgeLabs adds discovery, behavioural baselines, fleet posture, and audit-grade correlation across every host — turning a single-binary evaluator into a full agentic-security control plane.

A Host Layout AI EdgeLabs + Parallax
LLM PROVIDER Anthropic · OpenAI Local: Ollama, LM Studio AIEL CONTROL PLANE Fleet posture · Audit SIEM · Webhooks AI EDGELABS DEPLOYED ON HOST One agent container — protects 1..N AI agents & workloads on this host. eBPF · syscalls · pcap · DPI · < 4% CPU · zero egress · air-gapped HOST AI AGENT LLM client Anthropic / OpenAI / local SDK Tool runtime exec · file · http · shell · custom Memory & context conversation · RAG · state FRAMEWORKS Claude Code · OpenClaw LangChain · CrewAI · Agents SDK PARALLAX · RUST · < 0.2 MS Lifecycle Hooks message.before · tool.before · tool.after · params.before Evaluator chain — 5 engines regex → pattern → sigma → CEL → SQL cost-ordered · short-circuit on first block Decision engine block · redact · detect · allow Audit + Webhook JSONL · POST → SIEM AIEL RUNTIME AGENT eBPF probes syscalls · process tree Network capture pcap · DPDK · DPI File & vuln scan SBOM · drift · YARA Detection engine ML · APT rules · correlator /evaluate decision proxy mode
5engines

Cost-ordered evaluator chain

Regex, keyword pattern, Sigma rules, CEL expressions, and SQL-based temporal analysis run in cost order, cheapest first. The chain short-circuits the moment any rule returns block — layered defense without paying for it on every call.

4stages

Lifecycle interception

Hook into message.before, tool.before, tool.after, and params.before. Every prompt the model sees, every tool it tries to call, every result returned, and every parameter forwarded — inspected before execution, every time.

< 0.2ms

Microsecond decisions

Single static Rust binary, zero runtime dependencies — no Python, no JVM, no containers required. Typical decisions complete in under 0.2 ms; agents don't feel the inspection, attackers don't get the window.

2modes

Integrate any way you ship

Server mode exposes POST /evaluate for any framework that speaks HTTP. Proxy mode drops in front of Anthropic or OpenAI APIs and evaluates every request, response, and streaming tool call — with zero code changes.

Rollout — five stages from deploy to compliance

Stage 01
Deploy AIEL agent
Single container on each host — runtime visibility into processes, network, and files from day one.
Stage 02
Attach Parallax
Native plugin, lifecycle hooks, or transparent proxy between the agent and its LLM provider.
Stage 03
Evaluate every call
Each message and tool call runs the evaluator chain in under 0.2 ms — 51 rules across 13 threat categories out of the box.
Stage 04
Block or redact
Destructive commands and prompt-injection chains blocked, secrets and PII redacted — before they leave the host.
Stage 05
Correlate & report
Decisions stream into the EdgeLabs fleet posture, audit log, and SIEM — alongside every other host-level event.
The Problems

Why AI agents need external security

LLM-powered agents introduce an entirely new risk surface. Language is non-deterministic. Guardrails are suggestions. And agents operate faster than any human can observe.

Guardrails are just recommendations

Execution guardrails inserted into LLM configs are soft suggestions — models can and do ignore them. Just 9 seconds were needed to delete a production database. External enforcement is the only way to guarantee compliance.

Non-deterministic execution

Every agent execution can produce different results. Language is non-deterministic, and so is the outcome. Teams need monitoring for suspicious divergence from the agent's standard operational scope.

Limited observability

Agents make decisions and execute fast — there's no room for humans to observe it all. Organizations need a clear view into agent operations: domains contacted, processes spawned, resources consumed.

Shadow AI proliferation

AI agents can ease everyday operations but introduce decentralized risk — separate locations of sensitive information and vulnerabilities outside traditional CTO/CISO/COO department structures.

Fleet Posture Dashboard

Real-time fleet posture for every AI agent

See every agent across your fleet at a glance — which frameworks they run (OpenClaw, LangChain, CrewAI, OpenAI Agents, Anthropic SDK), their protection state, risk score, active rule coverage, and the latest blocks and detections.

  • Blocked attacks, redacted egress, and suspicious events — in real time
  • Latest blocks feed with severity, rule ID, and description
  • Top threat categories with 24-hour trending (prompt injection, dangerous commands, data exfiltration, PII, reconnaissance)
  • Per-agent risk score, rule coverage, and gap analysis
AI/LLM Security fleet posture dashboard showing blocked attacks, latest detections, top threat categories, and per-agent status Fleet Posture
Defense-Layer Coverage

Defense coverage across the fleet

Each row shows which Parallax evaluators are active per agent. Instantly spot missing coverage and surface it as a vulnerability — before attackers do.

  • Evaluator types: prompt-injection guard, dangerous commands, privilege escalation, PII scanner, secret scanner, rate limits
  • Hook-point visibility: message.before, tool.before, tool.after
  • Status indicators: nominal, high load / risk, critical anomaly
  • Rule coverage and gap counts per agent
AI/LLM Security defense-layer coverage showing per-agent evaluator status, hook points, and gap analysis Defense Coverage
Threat Categories

51 rules across 13 categories

Parallax evaluates every tool call against a comprehensive, configurable rule set. All rules ship out of the box and are fully customizable.

Threat Example Action
Destructive commands rm -rf /, mkfs, dd Block
Privilege escalation sudo, chmod u+s Block
Secret exfiltration AWS keys, GitHub PATs Redact
Prompt injection "ignore previous instructions" Block
Reconnaissance .aws/credentials, /etc/shadow Block
Supply chain attacks pip --index-url, curl | bash Block
PII leakage SSNs, credit card numbers Redact
Model manipulation Temperature overrides Block
Data exfiltration Bulk data export, unauthorized API calls Block
Unauthorized network access New outbound domains, reverse shells Block
File system tampering Config overwrites, log deletion Block
Resource abuse Crypto mining, GPU hijacking Block
Compliance violation Unapproved data processing regions Block

See what your AI agents are really doing — and enforce the rules that matter.

Deploy external guardrails, detect shadow AI, and prevent data loss across your AI fleet. One platform. Full runtime visibility. Sub-millisecond enforcement.