The AI agent security lifecycle.
Open source tools covering build, scan, protect, verify, monitor, and train. Each tool runs standalone. The OpenA2A CLI ties them together.
$ npx opena2a-cli review30 second security assessment. No account required.
Three pillars. Six stages.
Discover risk on your agent surface. Defend credentials, access, and runtime behavior. Validate trust and benchmark compliance.
Identify risk across your AI agent surface.
Protect credentials, access, and runtime behavior.
Verify trust, benchmark compliance, train teams.
Start here.
Pick the path that fits how you work today. Every tool is open source and runs locally.
I am building an agent
Run a full security review. Credential scan, config integrity, Shield posture, and deep scan in one command.
$ npx opena2a-cli reviewI am auditing agents
318 static checks plus 164 adversarial payloads. OASB compliance benchmark and auto remediation.
$ npx hackmyagent secureI am evaluating agent identity
How OpenA2A AIM (Agent Identity Management) registers agents, checks their actions, scores them and records MCP attestations, each statement linked to its source.
Read the AIM architectureAll tools.
Each tool works standalone. The OpenA2A CLI ties them together through a unified interface.
OpenA2A CLI
Build
Unified entry point that orchestrates every OpenA2A tool through adapter commands.
$ npx opena2a-cli review- Trust score 0 to 100
- Scope drift detection
- Config integrity signing
- Shadow AI discovery
HackMyAgent
Scan
318 static checks across 73 categories, 29 NanoMind semantic checks across 7 analyzers, and 164 adversarial payloads across 16 categories, with auto remediation and rollback.
$ npx hackmyagent secure- 318 static checks across 73 categories
- 29 NanoMind semantic checks across 7 analyzers
- 164 adversarial payloads across 16 categories
- Auto remediation with rollback
- OASB compliance reports
Secretless AI
Protect
Stores API keys outside project files and sets up AI coding tools not to read credential files. In Claude Code a PreToolUse hook blocks credential file reads. For Cursor, Copilot, and Devin Desktop (formerly Windsurf) it writes an instruction file that nothing enforces.
$ npx secretless-ai init- 57 credential patterns
- Multi backend storage
- MCP server encryption
- Transcript DLP scanning
AIM
Build
Cryptographic identity, capability policies, trust scoring, and audit logging for AI agents.
$ docker compose up- Ed25519 identity
- Policy as code
- 9 factor trust score
- MCP attestation
OASB
Scan
223 attack scenarios across 10 test categories, mapped to 15 MITRE ATLAS techniques. OASB-1 standard: 46 controls in 10 categories at 3 maturity levels.
$ npx opena2a-cli benchmark- 223 attack scenarios
- 10 test categories
- 15 MITRE ATLAS techniques
- Compliance reports
Runtime Protection
Monitor
Runtime process, network, and filesystem monitoring with protocol aware AI endpoint detection.
$ npx opena2a-cli runtime start- Process monitoring
- Network inspection
- MCP, A2A, and OpenAI detection
- Filesystem watching
AI Browser Guard
Protect
Chrome extension that detects and monitors AI agents in your browser, blocks scripted navigations, form submissions, and downloads best-effort for in-page automation, and closes agent tabs with a kill switch.
$ Install from Chrome Web Store- 3 layer agent detection
- Kill switch closes agent tabs
- Delegation wizard
- Session timeline
DVAA
Train
Deliberately vulnerable AI agent platform with 21 agents, 22 CTF challenges, and 3 protocol types.
$ docker pull opena2a/dvaa- 21 vulnerable agents
- 12 vulnerability categories
- 22 CTF challenges
- MCP, A2A, and OpenAI
ai-trust
Verify
Trust verification CLI for AI packages. Queries OpenA2A community trust data for security scans, dependency risk, and known advisories.
$ npx ai-trust check <package>- Single package trust lookup
- Dependency audit
- Batch verification
- 5 trust levels
- Scan on demand
- Community contributions
aicomply
Protect
Inline classifier for agent I/O. Detects PII, credentials, regulated data, and prompt-injection attempts before content reaches a hosted LLM.
$ npx @opena2a/aicomply scan- PII, credential, and regulated-data detection
- Deterministic regex plus semantic layer
- Masked output and CI exit codes
- npm and PyPI packages
Coverage matrix.
Map your threat model to the right tools. Every row is covered by at least two tools.
| Threat | CLI | HMA | Secretless | AIM | OASB | Runtime | ai-trust | aicomply | Guard | DVAA |
|---|---|---|---|---|---|---|---|---|---|---|
| Hardcoded credentials | . | . | . | . | . | . | ||||
| Prompt injection | . | . | . | . | . | . | ||||
| MCP config tampering | . | . | . | . | . | . | ||||
| Scope drift | . | . | . | . | . | . | . | . | ||
| Unauthorized agent access | . | . | . | . | . | . | . | |||
| Supply chain risk | . | . | . | . | . | . | . | |||
| Data exfiltration | . | . | . | . | . | . | ||||
| Privilege escalation | . | . | . | . | . | . | . |
Standards and governance.
Tools alone do not make agents safe. The OpenA2A specifications, the SOUL.md governance file, and the Open Agent Security Benchmark give you a shared vocabulary to design, score, and verify agent behavior.
SOUL.md
Behavioral governance file. Tier aware profiles for conversational, code assistant, tool agent, autonomous, and orchestrator agents.
OASB
Open Agent Security Benchmark. 223 attack scenarios across 10 test categories and 15 MITRE ATLAS techniques.
OpenA2A specifications
The open specifications for agent identity, trust and authorization, grouped by how far along each one is.
One command. Every tool.
The OpenA2A CLI version 0.10.13 delegates to HackMyAgent 0.33.2, Secretless AI, AIM, and ai-trust through a single interface. Pick a tool above or run the unified review.
$ npx opena2a-cli review