Feb 19|ARP v0.2.0 -- AI-layer interceptors, HTTP proxy, 19 detection patterns
All updates

Open Agent Security Benchmark

The standard for
AI agent security

Three measurement systems. OASB-1 checks infrastructure security — 46 controls, 3 maturity levels. OASB-2 checks behavioral governance — 72 controls, 9 domains, 4 agent tiers. OASB Eval tests whether security products detect real attacks — 182 scenarios, 10 MITRE ATLAS techniques.

$npx hackmyagent secure --benchmark oasb-1

Three measurement systems

One benchmark, three specifications

OASB-1Compliance

Check agent compliance

CIS Benchmarks for AI agents. 46 controls across 10 categories with L1/L2/L3 maturity levels. Answers: “Is your agent secure?”

Tests46 controls, L1/L2/L3
Analogous toCIS Benchmarks
AudienceAgent developers, compliance teams
View OASB-1 specification
OASB-2Governance

Govern agent behavior

Behavioral governance for AI agents. 72 controls across 9 domains with 4 agent tiers. Answers: “Does your agent behave correctly?”

Tests72 controls, 4 tiers
Analogous toSOC 2 Trust Principles
AudienceAgent builders, governance teams
View OASB-2 specification
OASB EvalEvaluation

Evaluate security products

MITRE ATT&CK Evaluations for AI agent security products. 182 attack scenarios across 10 MITRE ATLAS techniques. Answers: “Does your EDR catch this?”

Tests182 attack scenarios
Analogous toMITRE ATT&CK Evaluations
AudienceSecurity product vendors, evaluators
Explore OASB Eval

Protects against

Common AI agent vulnerabilities

Prompt injection

Malicious inputs that override system instructions, causing agents to ignore safety rules or execute unintended actions.

"Ignore previous instructions and..."
Protected by3.13.2

Jailbreaking

Techniques that bypass model guardrails through roleplay, encoding tricks, or adversarial prompts.

DAN, base64 encoding, character roleplay
Protected by3.33.1

Data exfiltration

Extracting sensitive information through crafted queries, side channels, or manipulated tool outputs.

Embedding data in URLs, markdown injection
Protected by4.34.1

Credential theft

Exposing API keys, tokens, or secrets through prompt leaks, logs, or unprotected storage.

System prompt extraction, log scraping
Protected by5.15.2

Tool abuse

Misusing agent capabilities beyond intended scope - file access, network calls, or code execution.

Path traversal, SSRF, command injection
Protected by2.12.27.1

Context poisoning

Manipulating conversation history or RAG sources to influence future agent behavior.

Memory injection, document poisoning
Protected by8.18.2

Internet-wide scan data

The current state of AI agent security

HackMyAgent scanned the public internet for exposed AI agent infrastructure. The results informed which OASB controls matter most.

97,127

Hosts discovered

11,192

Hosts scanned

1,594

Vulnerable

1,190

CLAUDE.md exposed

645

MCP tools exposed

5,042

Outdated endpoints

Read the full research report

Maturity levels

Choose your security posture

L1
Essential14 controls

Baseline security every agent should have. Development and prototypes.

Cryptographic identity
Prompt injection defense
Secret management
Execution isolation
L2
Standard11 controls

Production agents handling sensitive data. Most deployments.

Audit logging
Dynamic permissions
Dependency scanning
Context verification
L3
Hardened7 controls

Regulated industries and high-security environments.

Hardware isolation
Zero standing privileges
Anomaly detection
Automated response

Security controls

46 controls across 10 categories

ID
Control
Category
1.1
Agent Cryptographic Identity
Identity
1.2
Identity Attestation
Identity
2.1
Least Privilege Enforcement
Authorization
2.2
Permission Scope Validation
Authorization
3.1
Prompt Injection Defense
Input
3.2
Boundary Enforcement
Input
3.4
URL and Resource Validation
Input
4.1
Response Filtering
Output
4.3
Data Exfiltration Prevention
Output
5.1
No Hardcoded Secrets
Credentials
5.2
Secure Secret Storage
Credentials
6.1
Verified Component Sources
Supply Chain
6.2
Cryptographic Integrity
Supply Chain
7.1
Execution Isolation
Isolation

Open-source toolkit

Meet the requirements with open-source tools

Every OASB control maps to a free, open-source tool. Scan, fix, and verify compliance without vendor lock-in.

OpenA2A CLI

Unified security CLI that orchestrates all OpenA2A tools. Scan, protect, guard, and verify agents from a single command.

All 10 categories
$npx opena2a
View on GitHub →

HackMyAgent

Security scanner and red team toolkit. 147 checks, 55 attack payloads, auto-fix with rollback. Includes ARP runtime protection and OASB benchmarking.

3. Input Security4. Output Security6. Supply Chain8. Memory9. OperationsRuntime
$npx hackmyagent secure
View on GitHub →

Secretless AI

Keeps secrets out of AI context windows. PreToolUse hooks block credential access across Claude Code, Cursor, and Copilot.

5. Credential Protection
$npx secretless-ai init
View on GitHub →

AIM

Agent Identity Management -- cryptographic identity, capability enforcement, audit logging, and trust scoring for AI agents.

1. Identity2. Authorization7. Agent-to-Agent10. Monitoring
$pip install aim-sdk
View on GitHub →

Browser Guard

Chrome extension that detects and controls AI agents operating in the browser. Four-layer detection, delegation engine, and session timeline.

DetectionBrowser Security
$Chrome Web Store
View on GitHub →

DVAA

Damn Vulnerable AI Agent -- 10 intentionally vulnerable agents for training. Practice all 10 OASB categories in a safe environment.

All 10 categories
$docker run -p 3000:3000 opena2a/damn-vulnerable-ai-agent
View on GitHub →

Registry

Agent trust registry for the OpenA2A ecosystem. Discover, verify, and track trust scores for AI agents and security tools.

6. Supply Chain7. Agent-to-Agent
$registry.opena2a.org
View on GitHub →

Verify your agent's security

Run the benchmark against your AI agent. Read the docs for CI/CD integration.

$npx hackmyagent secure --benchmark oasb-1