Skip to content

severity-calibration-rubric.md

Latest commit

 

History

History
27 lines (19 loc) · 2.76 KB

severity-calibration-rubric.md

File metadata and controls

27 lines (19 loc) · 2.76 KB
Raw

Security Severity Calibration Rubric

This document defines the shared security severity scale for all vulnerability disclosures, threat models, risk registers, and module documentation in the TeaAgent repository.

Severity Scale

The severity of a security gap, threat, or vulnerability is determined by the potential impact of exploitation on the TeaAgent trust boundary:

Severity Definition Exploit/Bypass Impact Remediated Cadence / Gate
Critical Complete trust boundary bypass. Bypasses approval gates entirely, allows forgeability of the audit trail, executes arbitrary commands directly on the host machine (no container isolation), or causes irreversible data deletion. Immediately Blocking. Must be resolved before commit/push is declared complete.
High Severe compromise of security defense-in-depth or resource controls. Exposes sensitive host credentials/files, bypasses cost/budget boundaries (causing financial risk), spawns Docker containers without isolation flags (allowing network access), or expands path-restricted approvals to global access. Blocker for release packaging. Must be addressed within the current development sprint.
Medium Forensic, validation, or minor trust enforcement flaws. Replay window for authentication tokens, partial undo visibility, silenced warnings on storage failure, or absence of sandbox runtime warning flags. Should be addressed within the active development cycle.
Low Minor usability, UI inconsistency, or code cleanliness issues. Stale metadata properties, cosmetic logging errors, or UI cost display desynchronization. Backlog items. Checked during regular maintenance passes.

Calibrated Severity Mapping

Based on the shared rubric, the core security mechanisms map to the following calibrated levels:

  1. DANGER_FULL_ACCESS Mode: Critical (Bypasses all checks and gates by design; entering this mode must be cryptographically signed and recorded in the hash-chained audit log).
  2. allow_all_destructive Bypass: Critical (If allowed to bypass approval without explicit confirmation, it constitutes a total security boundary bypass).
  3. Audit Chain Forgeability: Critical (Ephemeral HMAC keys allow attackers to reconstruct the SHA-256 hash chain, defeating non-repudiation).
  4. Subagent Process Isolation: High (Running subagents on host or using root-owned Docker containers with network access enables filesystem escape or data exfiltration).
  5. Path-Scoped Expansion: High (Empty paths expanding to global workspace access violates the principle of least privilege).
  6. Credential Leakage via Inspect: High (Treating mutating commands like cat/head as read-only inspection allows exfiltration of keys/secrets).