Open Agent Security Benchmark
MITRE ATT&CK Evaluations for AI Security Products

222 standardized attack scenarios that evaluate whether a runtime security product can detect and respond to threats against AI agents. Mapped to MITRE ATLAS and OWASP Agentic Top 10.

terminal
$ npx @opena2a/oasb run

  OASB v1.0 -- Open Agent Security Benchmark

  Running:
    Atomic tests .......... 65/65
    Integration tests ..... 8/8
    Baseline tests ....... 3/3
    E2E tests ............ 6/6

  Results:
    Detected: 71/82
    Missed: 8/82
    False positives: 0/3

  Coverage: 86.6% detection rate

What Is OASB

OASB evaluates security products, not agents. It answers a specific question: can your runtime security tool detect and block attacks against AI agents? This is the same concept as MITRE ATT&CK Evaluations, which test endpoint security products against known adversary techniques -- applied to AI.

 OASBHackMyAgent
PurposeEvaluate security productsPentest agents
AnalogyMITRE ATT&CK EvaluationsOWASP ZAP
TargetRuntime security toolsAI agents themselves
OutputDetection rate scorecardVulnerability report

Test Categories

Four categories of tests that cover discrete detection, multi-step chains, false positive validation, and real OS-level execution.

Atomic Tests

65

Discrete detection tests covering OS-level system calls and AI-layer attacks. Each test isolates a single technique for precise evaluation.

Integration Tests

8

Multi-step attack chains that combine techniques into realistic scenarios. Tests whether products detect coordinated threats.

Baseline Tests

3

False positive validation using benign operations. Ensures security products do not block legitimate agent behavior.

E2E Tests

6

Real OS-level detection tests that execute actual system operations. Validates runtime interception capabilities.

MITRE ATLAS Coverage

Every test scenario maps to a MITRE ATLAS technique. OASB covers 10 techniques across the adversarial ML threat landscape.

Technique IDTechnique Name
AML.T0046Unsafe ML Inference
AML.T0057Data Leakage
AML.T0024Exfiltration
AML.T0018Persistence
AML.T0029Denial of Service
AML.T0015Evasion
AML.T0054Jailbreak
AML.T0056MCP Compromise
AML.T0051Prompt Injection
AML.TA0006Defense Response

AI-Layer Tests

40 tests targeting the AI-specific attack surface: prompt input and output scanning, MCP tool call validation, and A2A message inspection.

Prompt Input Scanning

14 tests

Tests whether the product detects malicious instructions embedded in user prompts, system prompts, and injected context.

Prompt Output Scanning

12 tests

Tests whether the product detects sensitive data, credential leakage, and unsafe content in model outputs.

MCP Tool Call Validation

8 tests

Tests whether the product validates tool calls for parameter injection, unauthorized access, and privilege escalation.

A2A Message Scanning

6 tests

Tests whether the product inspects inter-agent messages for instruction injection, data exfiltration, and trust boundary violations.

Quick Start

Clone the repository, install dependencies, and run the benchmark against your security product.

terminal
# Clone the repository
$ git clone https://github.com/opena2a-org/oasb.git
$ cd oasb

# Install dependencies
$ npm install

# Run all tests
$ npx @opena2a/oasb run

# Run a specific category
$ npx @opena2a/oasb run --category atomic

# Run a specific MITRE technique
$ npx @opena2a/oasb run --technique AML.T0051
222
Attack scenarios
10
MITRE techniques
82
Test files
4
Categories

Every OASB control has a tool

The OpenA2A ecosystem covers all 10 OASB categories. Each control maps to one or more open-source tools.

Identity & Provenance

AIM

Ed25519 + ML-DSA post-quantum keypairs, ownership registry, agent bill of materials

Capability & Authorization

AIM

Capability-based access control, JIT access grants, runtime enforcement

Input Security

HackMyAgentARP

75 attack payloads across 7 categories, runtime prompt interception

Output Security

HackMyAgentARP

Output validation, exfiltration detection, runtime output scanning

Credential Protection

Secretless AIHackMyAgent

49 credential patterns, MCP vault protection, context window isolation

Supply Chain Integrity

HackMyAgentRegistry

SkillGuard hash pinning, SignCrypt signing, registry trust verification

Agent-to-Agent Security

AIMHackMyAgent

Mutual authentication, 10 A2A attack payloads, trust boundaries

Memory & Context

HackMyAgentARP

Context manipulation testing, runtime memory isolation

Operational Security

HackMyAgentARP

147 configuration checks, process/network/filesystem monitoring

Monitoring & Response

AIMARP

8-factor trust scoring, behavioral anomaly detection, kill switch

Benchmark your security product

npx @opena2a/oasb run