Please do not report security vulnerabilities through public GitHub issues.

Arbitus is a security-focused project. We take all security issues seriously and appreciate responsible disclosure.

Report security vulnerabilities via one of these methods:

1. GitHub Security Advisories (Preferred)

   • Navigate to github.com/arbitusgateway/arbitus/security/advisories
   • Click "Report a vulnerability"
   • Provide detailed description of the vulnerability

2. Email (Once configured)

   • Send to: security@arbitusgateway.dev
   • Include: "SECURITY: Arbitus Vulnerability" in subject line
   • Provide: Affected version, reproduction steps, potential impact

Please provide the following information:

• Description of the vulnerability
• Steps to reproduce the issue
• Affected versions (if known)
• Potential impact and attack scenario
• Proof of concept (if available)
• Suggested fix (if you have one)

1. Acknowledgment: We confirm receipt within 24 hours
2. Triage: We assess severity and assign a CVE if applicable
3. Development: We develop a fix privately
4. Review: Security team reviews the fix
5. Release: We release the fix and publish an advisory
6. Disclosure: We credit the reporter (if desired) in the advisory

Security updates are provided for the following versions:

We recommend always running the latest stable release.

Arbitus includes the following security features:

• Per-agent policies: API keys, JWT/OIDC, mTLS
• Tool allowlists/denylists: Wildcard patterns supported
• Resource access control: allowed_resources/denied_resources
• Prompt access control: allowed_prompts/denied_prompts
• OPA/Rego policies: Custom policy evaluation

• Payload filtering: Block or redact sensitive patterns
• Encoding-aware: Base64, URL-encoded, Unicode variants
• Schema validation: Validate against inputSchema
• Prompt injection detection: Built-in heuristics

• Per-agent limits: Sliding window enforcement
• Per-tool limits: Tool-specific rate limits
• IP-based limits: Infrastructure-level protection

• Immutable audit log: SQLite with hash-chain integrity
• Fan-out: Multiple backends simultaneously
• CloudEvents 1.0: SIEM-compatible format
• OpenLineage: Lineage tracking

• Binary verification: SHA-256 hash pinning
• Cosign verification: Sigstore bundle verification

• Approval workflow: Sensitive operations require human approval
• Shadow mode: Dry-run risky operations

• HTTP transport: Use TLS in production
• mTLS available: For agent authentication
• Stdio transport: Runs locally; less exposure

• Secrets management: Use ${VAR} interpolation or Vault/OpenBao
• Hot reload: Changes apply without restart; protect config files
• Admin endpoints: Require Authorization: Bearer <admin_token>

• Circuit breaker: Protects against upstream failures
• Timeout enforcement: Prevents indefinite hangs
• Response filtering: Blocks sensitive patterns in upstream responses

We recognize security researchers who responsibly disclose vulnerabilities.

(None yet - be the first!)

• Security Issues: See reporting instructions above
• General Questions: Open a GitHub Discussion
• Maintainer Contact: See MAINTAINERS.md

Thank you for helping keep Arbitus and its users secure!