AgentSentry
NHI · AI AGENT SECURITY
Open source · v0.1.0

Find every Machine's Identity before they do.

AI agents and service accounts are the new attack surface — ungoverned, under-monitored, and multiplying fast. AgentSentry discovers every non-human identity in your cloud, scores each one with a provable risk formula, and gives you a fix-it checklist before attackers find what you haven't.

1,610+
Active CVEs
45:1
Machine/human ratio
Open Source
MIT Licensed
agentsentry — static agent audit

Target repository

sre-copilot/
agents/
agent_executor.py
tools/
execute_code.py
shell_tool.py
http_tool.py
chains/
rag_chain.py
prompts/
.env
requirements.txt
langgraph.json

user@sec-workstation:~/sre-copilot$

0+
CVEs in CISA KEV catalog
Updated daily. Free.
0:1
Machine to human identity ratio
Almost none governed.
0+
KEV entries linked to ransomware
Active campaigns. Right now.
How it works

From zero to attack graph
in under three minutes.

No agents to deploy. No SaaS data upload. Runs entirely local — your cloud credentials never leave your machine.

01

Discover

Point AgentSentry at your AWS account. It enumerates every IAM role, access key, service account, OAuth token, and AI agent in minutes — including ones you forgot existed.

02

Score

Each identity gets a P×R×E×A risk score: Privilege × Reachability × Exposure × AI-Amplification. Critical identities surface immediately. CISA KEV enrichment flags active CVEs.

03

Visualize

An interactive attack graph shows every identity and the access paths between them. See exactly what an attacker could reach if any given identity is compromised.

Providers

Not just AWS.
Everywhere you deploy.

Six independent providers — install only what you need. Each one checks its own permissions before touching a single API. Start with local — it needs nothing and finds more than you expect.

Amazon Web Services
agentsentry scan aws
  • IAM Roles & Access Keys
  • Lambda execution roles
  • S3, RDS, Secrets Manager
$ pip install agentsentry[aws]
$ aws configure
Microsoft Azure
agentsentry scan azure
  • Managed Identities
  • Service Principals
  • Role assignments (Owner/Contributor)
$ pip install agentsentry[azure]
$ az login
Google Cloud
agentsentry scan gcp
  • Service Accounts
  • User-managed SA keys
  • Project IAM bindings
$ pip install agentsentry[gcp]
$ gcloud auth application-default login
GitHub
agentsentry scan github
  • Personal Access Tokens
  • Deploy Keys & SSH Keys
  • Actions Secrets
$ export GITHUB_TOKEN=<pat>
Kubernetes
agentsentry scan k8s
  • ServiceAccounts & RBAC
  • ClusterRoleBindings
  • Automount token exposure
$ pip install agentsentry[k8s]
$ kubectl config use-context <cluster>
Local Environment
agentsentry scan local
No credentials needed
  • Env vars & .env files
  • SSH keys & credential files
  • Docker socket & git tokens
AUTO-DETECT & SCAN EVERYTHING
$agentsentry scan all
Detects which providers are configured and scans them all in one pass.

$ agentsentry blast nhi://agent/prod-sre-copilot

select a node

sts:AssumeRoles3:GetObjectlambda:InvokeGetSecretValueProduction SRE CopilotLangChain agent · L4 autonomysre-copilot-execIAM role · trust policy: *customer-data-prodS3 bucket · PII, billingreport-rendererLambda · same trust scopeprod/db-credentialsSecrets Manager
compromised / impacted exposed in blast scope
blast_radius.json
{  "source": "nhi://agent/prod-sre-copilot",  "compromise_assumed": true,  "hops": [    {      "via": "iam_role",      "id": "sre-copilot-exec",      "edge": "sts:AssumeRole"    },    {      "via": "policy",      "id": "AmazonS3FullAccess",      "edge": "s3:GetObject"    }  ],  "terminal_asset": "arn:aws:s3:::customer-data-prod",  "assets_reachable": 14,  "data_classes": [    "PII",    "billing"  ],  "prea": {    "P": 5,    "R": 2.5,    "E": 3,    "A": 50,    "score": 1875,    "severity": "CRITICAL"  }}

14 assets reachable from one compromised agent · PREA 1875 [CRITICAL]

What it does

Every attack surface.
One scanner.

The only open-source tool that audits machine identities across every cloud and environment — with the same risk model, in the same scan.

Multi-Cloud NHI Discovery

Finds every IAM role, API key, service account, Managed Identity, and OAuth token — across AWS, Azure, GCP, GitHub, Kubernetes, and your local machine. One command. Every environment.

AI Agent Scanner

Statically analyzes LangChain, CrewAI, and AutoGen codebases. Extracts tool permissions. Computes the AI-Amplification Factor.

CISA KEV Enrichment

Correlates every finding against 1,610+ actively exploited CVEs. Flags ransomware-linked vulnerabilities in real time.

Attack Graph

Cross-provider attack graph. Computes blast radius: if this identity is compromised, what does the attacker reach — regardless of which cloud it lives in?

MITRE ATT&CK Mapping

Every finding maps to ATT&CK techniques. T1078.004, T1528, T1552, T1611 — the language your SOC already speaks.

Risk Scoring: P×R×E×A

Privilege × Reachability × Exposure × AI-Amplification. Consistent across all providers — the same score model whether the identity lives in AWS, K8s, or a local .env file. Novel academic contribution.

P×R×E×A Calculator

Compute any NHI's risk score.

Drag the sliders or pick a preset to see how Privilege, Reachability, Exposure, and AI-Amplification combine into a real risk score.

PPrivilege (0–10)
7
RReachability (0–10)
6
EExposure (0–5)
3
AAI-Amplification (1–3×)
1.5
7×6×3×1.5=189.0
CRITICAL
189.0
threshold: >100 CRITICAL · >50 HIGH · >20 MEDIUM
Real-world presets

§ 4 · Scoring methodology

The PREA risk model

Every non-human identity is scored by a single multiplicative model. The formulation is deliberately auditable: each factor is computed from statically observable evidence, and the final score decomposes into the exact terms shown in scan output.

Definition 4.1 — composite risk

Risk(n) = P(n) × R(n) × E(n) × A(n)

for each identity n ∈ N, the set of discovered non-human identities

PREA factor definitions and ranges
TermRangeMeasures
PPrivilege[0.1, 5.0]Effective permission scope of the identity's credentials
RReachability[0.1, 3.0]Network and trust-boundary exposure of the identity
EExposure[0.1, 3.0]Credential hygiene: rotation age, storage, plaintext leaks
AAI-Amplification[1.0, 50.0]Risk multiplier for autonomous agents with tool access

Example 4.1 — finding AS-0042

Risk = 5.0 × 2.5 × 3.0 × 50.0 = 1875.0 [CRITICAL]

The same identity with an L2 approval gate (A = 5.0) scores 187.5 — one configuration change moves the finding two severity bands.

§ 4.1–4.4 · The AI-Amplification factor

A is the term that distinguishes PREA from conventional identity-risk models. It isolates two statically measurable properties of an agent: what its tools can do irreversibly, and how unsupervised it is when doing it.

Each tool bound to an agent is classified by the reversibility of its worst-case invocation. Read-class tools (retrieval, search) carry a 1× weight. Write-class tools (file mutation, API POST) carry a 4× weight. Execute-class tools (code execution, shell, infrastructure mutation) carry a 10× weight, because a single invocation can establish persistence or destroy state with no rollback path. The classification is derived statically from the tool's declared schema and call sites — no runtime instrumentation is required.

Pricing

Free forever.
Pay once for Pro.

The scanner is open source and always will be. Pro unlocks reports, enrichment, and JSON output — one payment, lifetime license.

Community
$0
Open source · MIT · free forever
Get started free
  • AWS, Azure, GCP, GitHub & K8s scanners
  • LangChain / CrewAI / AutoGen agent scanner
  • P×R×E×A risk scoring engine
  • MITRE ATT&CK technique mapping
  • Blast radius analysis
  • CLI — runs locally, no data leaves you
  • Open source — MIT license
Available now
Pro
$49
One-time · license key · no subscription
Sign up to upgrade
  • Everything in Free
  • --visualize: interactive HTML attack graph
  • --enrich: CISA KEV threat intel enrichment
  • --json: JSON output for pipelines & CI
  • Interactive multi-cloud scan mode
  • One-time purchase — license key, yours forever
  • Priority email support
How it works:
1. Sign up free → activate with AS-FREE-XXXX
2. Upgrade from your dashboard → AS-PRO-XXXX

API key & activation code delivered instantly by email · CLI works offline

BLAST RADIUS BY AGENTSENTRY

Stay ahead of machine identity threats

Weekly intel on NHI security and AI agent risks — real findings, practical commands, no fluff. Join security engineers and DevOps teams already subscribed.

No spam. Unsubscribe anytime. Every Tuesday.