Now in Beta · OWASP Agentic Top 10 covered

The security kernel
for AI agents

Drop-in trust boundary for any agent system. Cryptographic identity, real-time policy enforcement, memory integrity, and automatic quarantine — in one SDK import.

Free trial · No credit card required · Cancel anytime

4

Hardened security agents

10/10

OWASP Agentic coverage

< 5ms

Policy check latency

256-bit

Ed25519 cryptography

The problem

AI agents are a new attack surface

Existing security tools were built for humans and APIs. They don't understand agent goals, tool scopes, or memory semantics.

⚠️

No identity between agents

Agents call each other with no cryptographic proof of who they are. Any compromised agent can impersonate any other.

🔓

No policy enforcement

Tool calls execute with no governance layer. A prompt injection attack can exfiltrate data, spawn processes, or delete records.

🧠

Memory is a silent attack surface

Poisoned vectors persist silently and corrupt future decisions. There's no anomaly detection on what agents write to memory.

How it works

Four hardened agents.
One trust boundary.

Guardian runs alongside your agents, not inside them. It intercepts, validates, and enforces — invisibly.

Identity Agent

LangGraph

Issues cryptographically signed passports per agent. Ed25519 signatures — forgery is computationally impossible, not just policy.

Ed25519 keypair per agent
Scoped capabilities list
Parent chain inheritance
Revocation in < 50ms

Governance Agent

CrewAI

Intercepts every tool call before execution. Semantic drift detection, param validation, privilege ceiling enforcement — in real time.

Goal hijack detection
Tool misuse blocking
Privilege escalation guard
Hot-reloadable policies

Memory Guardian

pgvector

Append-only vector store with anomaly detection on every write. Poison a memory vector and you get flagged before it spreads.

Cosine anomaly detection
Immutable append-only store
Snapshot rollback points
Namespace-level RLS

Containment Agent

LangGraph

Four escalation tiers from warn to terminate. Human override webhook. Quarantine freezes all pending tool calls instantly.

4-tier escalation ladder
Human override webhook
Passport revocation on quarantine
Full audit trail

Integration

One import.
Fully protected.

Guardian wraps your existing agents. No architectural changes, no new infrastructure, no rewrite. Import, register, check.

1
pip install guardian-sdk

Python and TypeScript SDKs

2
g = Guardian(api_key=...)

Connect with your API key

3
await g.governance.check(...)

Every tool call is now enforced

agent.py
#a5d6ff">"color:#ff7b72">from guardian "color:#ff7b72">import Guardian

g = Guardian(api_key=#a5d6ff">"gai_live_••••••••••••")

# Register your agent with scoped permissions
passport = #a5d6ff">"color:#ff7b72">await g.identity.register(
    agent_id=#a5d6ff">"research-agent-01",
    scopes=[#a5d6ff">"tools:web_search", "memory:read"],
)

# Every tool call now enforced by policy
#a5d6ff">"color:#ff7b72">await g.governance.check(
    passport=passport,
    tool=#a5d6ff">"web_search",
    params={#a5d6ff">"query": "quarterly report"},
    goal=#a5d6ff">"Summarize Q3 financials",
)
# → {"decision": "allow", "confidence": 0.97}

# Memory writes are anomaly-checked automatically
#a5d6ff">"color:#ff7b72">await g.memory.write(passport, vector=embedding, metadata={
    #a5d6ff">"source": "web", "timestamp": "2026-05-03"
})

Features

Everything you need.
Nothing you don't.

Zero-config integration

One import. Guardian wraps your existing agents — no architectural changes.

Full observability

Every decision logged to an immutable audit table with structured JSON.

Policy as code

Governance rules stored in Supabase, hot-reloaded every 30s without restarts.

Multi-tenant ready

Namespace isolation, scoped API keys, and per-org dashboards out of the box.

Crypto-bound identity

Passports are Ed25519 signed. Revoke in milliseconds, not minutes.

Real-time risk scoring

Per-agent risk scores updated continuously from governance decisions.

Security coverage

OWASP Agentic Top 10

Every category covered. Not just detected — actively blocked.

#AttackComponentHow
A01
Prompt Injection
GovernanceInput sanitization + pattern matching on every tool call
A02
Goal Hijacking
GovernanceSemantic cosine distance between registered goal and current action
A03
Privilege Escalation
IdentityScope ceiling — agents cannot request beyond parent's permissions
A04
Memory Poisoning
Memory GuardianAnomaly score on every vector write; flag + optional rollback
A05
Tool Misuse
GovernanceJSON Schema validation of tool call params against registered spec
A06
Data Exfiltration
Containment + RLSSupabase Row Level Security + rate limiting + quarantine on anomaly
A07
Agent Impersonation
IdentityEd25519 signature verification — DB lookup is not enough
A08
Resource Abuse
ContainmentPer-agent rate limit counters; auto-throttle before quarantine
A09
Supply Chain Attack
InfraPinned dependencies, SBOM, no dynamic imports at runtime
A10
Auth Bypass
Identity + RLSRevocation list checked on every request; RLS blocks stale tokens

Pricing

Simple, transparent pricing

Start free. Scale when you need to.

Starter

Free

For individual developers and small experiments.

5 active agents
10,000 governance checks / mo
1 API key
7-day audit log retention
Community support
Start free
Most popular

Pro

$49/ month

For teams shipping production AI agents.

50 active agents
500,000 governance checks / mo
10 API keys
90-day audit log retention
Real-time activity feed
Email support
Start free trial

Enterprise

Custom

Unlimited scale, SLA, dedicated support.

Unlimited agents
Unlimited checks
Custom policies
1-year log retention
99.9% SLA
Dedicated support
Contact us

Start protecting your agents today

Free tier includes 5 agents and 10,000 governance checks per month. No credit card, no commitment.