Originally published on opena2a.org

OASB: Why AI Agents Need CIS-Style Security Benchmarks

OpenA2A Team|
#oasb#benchmark#ai-agents#security#governance

AI agents are deploying to production faster than security teams can assess them. Most organizations have no structured way to evaluate whether an agent is safe to deploy. OASB (Open Agent Security Benchmark) brings the CIS Benchmark model to agentic AI -- providing a standardized framework with 46 controls, 10 categories, and 3 maturity levels.

The Standards Gap

Web applications have OWASP Top 10. Cloud infrastructure has CIS Benchmarks. Container deployments have the CIS Docker Benchmark. These frameworks share a common pattern: they define measurable controls, organize them into categories, and provide maturity levels so organizations can adopt them incrementally.

AI agents have none of this. They execute code, access filesystems, make network requests, manage credentials, and interact with other agents -- all without a standardized security baseline. Security teams are left writing ad-hoc checklists or adapting web application standards that do not account for agent-specific risks like prompt injection, tool misuse, or context poisoning.

What OASB Defines

OASB-1, the first specification, covers agent configuration security. It organizes 46 controls into 10 categories:

CategoryFocus
Identity & ProvenanceAgent identity verification and cryptographic signing
Capability & AuthorizationPermission boundaries and least privilege enforcement
Input SecurityPrompt injection protection and input validation
Output SecurityResponse sanitization and data leakage prevention
Credential ProtectionSecret management and credential hygiene
Supply Chain IntegrityDependency verification and package trust
Agent-to-Agent SecurityInter-agent authentication and message integrity
Memory & Context IntegrityContext injection protection and memory isolation
Operational SecurityResource limits, sandboxing, and process isolation
Monitoring & ResponseSecurity logging, alerting, and incident response

Three Maturity Levels

Not every agent deployment requires maximum hardening. OASB defines three levels that let organizations adopt controls incrementally:

L1Essential

Baseline security for any agent in production. Covers critical risks with minimal overhead -- credential hygiene, basic input validation, and identity verification.

L2Standard

Defense-in-depth for business-critical agents. Adds monitoring, supply chain verification, and enhanced access controls.

L3Hardened

Maximum security for regulated environments. Cryptographic verification, formal audit trails, advanced isolation, and full compliance mapping.

Automated Verification

OASB controls are designed to be machine-verifiable. HackMyAgent implements automated checks for every OASB-1 control:

# Run L1 benchmark against your project
$ npx hackmyagent secure --benchmark oasb-1

# Target a specific maturity level
$ npx hackmyagent secure -b oasb-1 --level L2

# Output as SARIF for CI/CD integration
$ npx hackmyagent secure -b oasb-1 -f sarif -o results.sarif

Each control check returns a pass/fail result with specific remediation guidance. Results map to SOC 2, ISO 27001, and NIST CSF for organizations that need compliance evidence.

Read the full post on opena2a.org, or explore the full specification at oasb.ai/controls.

OpenA2A is building open security infrastructure for AI agents. Follow our progress at opena2a.org.