Linux-first · Open-source security layer

Your AI can act. You decide how far.

MonkeyShield sits between an AI agent and your operating system, controlling what it can read, write, execute and access.

Explore the architecture View policy example

Deny by default Human approval Full audit trail

LIVE POLICY TRACE ACTIVE

AI AGENT local-agent-01 Qwen / Llama / custom

MONKEYSHIELD CORE Policy engine context · identity · intent · risk

ALLOWED file.read /workspace/docs/*

BLOCKED shell.execute sudo rm -rf /

01 / PROBLEM

AI agents are gaining tools faster than they are gaining boundaries.

The model is not the only attack surface. Every permission behind it multiplies the damage a bad instruction can cause.

Overprivileged agents

One process often receives broad access to files, shell commands, APIs and secrets just to complete a narrow task.

EXCESSIVE PERMISSIONS

Prompt injection becomes action

A malicious document or webpage can influence the agent. Without enforcement, text can become a system-level operation.

UNTRUSTED INPUT

Invisible decisions

When actions are scattered across tools and logs, it becomes difficult to answer who requested what, why and with which result.

LOW OBSERVABILITY

02 / CAPABILITIES

A security boundary designed around agent actions.

Deny by default

Nothing passes unless an explicit policy allows it.

↗

Human approval

Risky operations pause before execution and ask the operator.

↗

Tamper-evident audit

Every request, decision and policy version leaves a trace.

↗

Fine-grained scope

Control paths, commands, hosts, methods, identities and runtime context.

↗

monkeyshield://policy-engine

POLICY DECISION Default action: DENY

request: file.write
resource: /etc/ssh/sshd_config
identity: local-agent-01
matched_rule: none
result: BLOCKED

APPROVAL REQUIRED Operator confirmation requested

request: shell.execute
command: apt install tcpdump
risk_score: 72 / 100
mode: HUMAN_APPROVAL
timeout: 120s

AUDIT EVENT Action chained into signed log

event_id: ms_01JZP8V7
policy_hash: 6c0d…a918
previous_hash: 3ee4…0b11
integrity: VERIFIED

SCOPED ALLOW Least privilege in action

allow: file.read
paths:
  - /workspace/src/**
except:
  - **/.env
  - **/secrets/**
result: ALLOWED

03 / ARCHITECTURE

The agent never touches the system directly.

Each tool request is intercepted, enriched with context, evaluated against policy and only then executed—or stopped.

Click a technical term to see a plain-language explanation.

SOURCE

· ·

tool request

ENFORCEMENT

MonkeyShieldidentity · policy · approval · audit

controlled action

TARGET

read · write · delete

execute · process · privilege

host · method · destination

01 Detect 02 Evaluate 03 Approve 04 Enforce 05 Record

04 / POLICY AS CODE

Readable rules. Predictable behavior.

Security policies should be reviewable, versioned and easy to explain—not buried in opaque model behavior.

✓Path and command patterns ✓Per-agent identities and roles ✓Explicit approval thresholds ✓Versioned, auditable decisions

policy.yaml

version: "1"
default: deny

identities:
  local-agent-01:
    allow:
      - action: file.read
        paths:
          - "/workspace/**"
        except:
          - "**/.env"
          - "**/secrets/**"

      - action: shell.execute
        commands:
          - "git status"
          - "pytest *"
        approval_required: true

SECURITY SPECTRUM

Built where Red, Blue and Green become Purple.

Attack thinking finds the paths. Defensive engineering closes them. Safe operations keep the system useful. MonkeyShield combines all three.

RED

BLUE

GREEN

PURPLE MONKEYSHIELD

05 / ROADMAP

Start narrow. Build the boundary correctly.

NOW · MVP

Linux enforcement core

file.read / file.write
shell.execute
deny-by-default YAML policies
local audit trail

interactive approval broker
network.request controls
risk-based policy decisions
signed policy versions

LATER

Agent security platform

Windows support
RAG content quarantine
memory and secret protection
central observability

OPEN SOURCE · IN DEVELOPMENT

Give AI tools. Keep control.

MonkeyShield is being built in public by MonkeyPivot as the first project in the MonkeyAI Security Tools ecosystem.

Security model Open Playground↗

Repository link will appear here after the first public release.