AI Agent Security Posture
Your engineers are shipping AI agents faster than your security team can audit them. Amajoni discovers the agents running in your AWS account, maps each one's blast radius, and shows you the ones a single prompt injection could turn into an insider threat.
Actual scan output format — your results may vary
“I spent five years in SOC and pen testing watching AI agents get deployed into production with admin credentials and no security review. Nobody knew they were there. Nobody knew what they could do. That is the problem Amajoni solves.”
THATO DITSELE · FOUNDER & CEO, AMAJONI
The Problem
01
Every team ships agents with LangChain, CrewAI, n8n. Security has no inventory. The average enterprise has 5–10× more agents than they think.
02
Developers ask for the API key that works and get admin. That agent now has prod database access and shell execution. One prompt injection = game over.
03
Your SIEM doesn't know what anomalous tool use looks like. You'd never know an agent went rogue until the damage is invoiced.
82%
of enterprises have unknown or shadow AI agents running right now
CSA / Token
4%
of large organisations using AI agents have secured them
DigiCert
74%
of security leaders say their agents have more access than they need
8:1
agentic AI adoption is outpacing defences by roughly this much
Gartner
How It Works
Point Amajoni at your AWS account with read-only access. Nothing is installed, no policy is touched, and production traffic is never in the path.
Every agent found is scored against the IAM permissions it actually holds — not a canned risk label. An admin wildcard scores 100, and the breakdown names the permissions behind every number.
Findings map to NIST AI RMF, MITRE ATT&CK, OWASP LLM Top 10, EU AI Act, SOC 2 and POPIA — ready for your auditor, not just your dashboard.
Platform Capabilities
Finds Lambda, container and serverless workloads that reach an LLM — including ones nobody documented — from IAM permissions, environment config and runtime evidence. Classified Confirmed or Possible from that evidence, never from the name.
0–100, computed from the IAM permissions an agent actually holds. Every score carries a breakdown naming which permission contributed what — so you can argue with it.
A Lambda and the execution role it runs as are one agent, not two. Triage state survives across scans, and agents that stop appearing age out of the active inventory.
PDF evidence packs across NIST AI RMF, EU AI Act, POPIA, SOC 2, executive summary, vendor assessment and detection rules. Not Met is reported as Not Met.
Alerts deduplicated on a stable fingerprint, so the same finding twice is one alert with history. Slack notification on critical discoveries and a summary after every scan.
Passive OSINT scan of any public domain — finds AI agent signals in JavaScript bundles, API endpoints and response headers. No credentials needed.
Free Feature — No Credentials Needed
Enter any public domain and Amajoni finds AI agent signals in their JavaScript bundles, API endpoints, subdomains, and HTTP headers. No cloud access required.
→ JavaScript bundle analysis
→ API endpoint detection
→ Subdomain enumeration via crt.sh
→ HTTP header fingerprinting
Domain Intelligence Scan · Example
takealot.com
Confidence Score
74 / 100
HIGH PROBABILITY
JS Signals
Endpoints
/api/chat — 401 /api/assistant — 401
Example output · Enter a domain above to scan it
Findings map to
Positioning
Before you deploy runtime controls, you need to know what you're controlling. Amajoni discovers your AI agents, scores their risk, and generates the compliance evidence your auditor needs — read-only, with zero infrastructure changes.
It complements runtime enforcement tools rather than replacing them. A gateway or policy engine needs an inventory and a risk picture before it has anything to enforce. That inventory is what we produce.
And it is built POPIA-first. Most AI security tooling is designed around US and EU regimes and fits African data-residency and regulatory realities poorly, if at all. We started from the other end.
From our research
Using only passive OSINT — no credentials, no access — we found publicly visible AI infrastructure across the majority of South Africa's largest fintechs.
Read the full report →85%
had AI signals in public JS
55%
had confirmable AI API endpoints
20%
had LLM SDK in client-side code
A read-only connection and a scan. Nothing installed, no policy modified.