Skip to content

Latest commit

 

History

History
31 lines (23 loc) · 1.75 KB

File metadata and controls

31 lines (23 loc) · 1.75 KB

Zyra Threat Model

Security objective

Prevent an AI agent, plugin, skill, MCP-compatible integration, user, or compromised dependency from exceeding its declared authority or exposing protected data.

Trust boundaries

  • Browser and external API clients are untrusted.
  • Agent manifests and natural-language instructions are untrusted input.
  • Third-party models, packages, tools, webhooks, and remote content are untrusted until admitted.
  • Organization identity, policy decisions, and audit evidence remain inside the Zyra trust boundary.
  • Production credentials are never supplied to unapproved agent capabilities.

Priority threats and controls

Threat Primary controls
Prompt injection Manifest scanning, constrained tool schemas, deny-by-default policy
Tool or scope escalation Declared/requested scope comparison, RBAC, human approval
Secret exfiltration Pattern scanning, restricted-data egress denial, redacted evidence
Malicious agent skill Admission scanning, provenance, code review, sandboxing
Dependency compromise Lockfile, Dependabot, dependency review, CodeQL, SBOM/AI-BOM
Cross-tenant access Organization-scoped authentication and storage queries
Audit tampering Append-only application behavior and SHA-256 evidence hashes
Unsafe high-impact automation Review decision and owner/admin approval requirement
Weapon or physical-force use Explicit product boundary and policy denial

Residual risk

Static and behavioral scanners can produce false positives and false negatives. A clean scan is not proof that an agent or dependency is safe. Production changes still require human review, protected branches, least-privilege credentials, runtime isolation, and incident-response procedures.