Skip to content

Latest commit

 

History

History
60 lines (49 loc) · 2.4 KB

THREAT_MODEL.md

File metadata and controls

60 lines (49 loc) · 2.4 KB

Threat Model

Assets

  • Cloud credentials and short-lived tokens
  • SSH keys, Kubeconfigs, PAM credentials, and HSM references
  • Customer data and security telemetry
  • Agent prompts, tool definitions, memory, and policy files
  • Terraform state and plan artifacts
  • Evidence ledger and audit history
  • Skill artifacts and approval records
  • IP, policy models, connector code, and deployment topology

Threat actors

  • Malicious or compromised skill author
  • Typosquatted or dependency-confused package
  • Prompt-injection content in a skill, issue, webpage, or tool result
  • Compromised MCP server
  • Malicious insider or over-privileged integrator
  • Stolen workstation or phone
  • Compromised cloud account
  • Supply-chain attacker
  • Accidental operator error
  • External attacker reaching a management endpoint

Primary abuse cases

  1. Skill reads .env, SSH keys, cloud configuration, browser sessions, or Kubernetes secrets.
  2. Skill tells the agent to ignore its safety rules.
  3. Skill downloads and executes a remote script.
  4. Skill silently adds an MCP server with broad tools.
  5. Skill persists through launch agents, cron, systemd, or shell profiles.
  6. Agent applies a destructive Terraform plan without approval.
  7. Compromised connector mints a broader credential than requested.
  8. RDP or SSH endpoint is exposed publicly.
  9. Customer evidence is mixed across tenants.
  10. A provider update changes behavior without a re-scan.

Mitigations

  • SkillSpector pre-install and pre-upgrade gate
  • Immutable artifact pinning and approval ledger
  • Quarantine and non-executing static scan
  • Disposable execution runners
  • Network, filesystem, secret, and process allowlists
  • Short-lived capability credentials
  • Policy decision before every tool call
  • Human and second-factor approval for high-risk mutation
  • Append-only signed evidence
  • Tenant isolation and customer-owned data boundaries
  • Tailscale/private overlay instead of public management ports
  • Separate development, staging, production, and regulated enclaves
  • External review and periodic credential rotation
  • Incident response, revocation, and rollback procedures

Residual risk

SkillSpector is static analysis and optional semantic analysis. It is not malware detonation, host sandboxing, or proof of benign intent. Binary, encrypted, time-delayed, or novel attacks may evade it. Runtime isolation, least privilege, provenance, and human approval remain mandatory.