Skip to content

THREAT_MODEL.md

Latest commit

 

History

History
119 lines (80 loc) · 3.15 KB

THREAT_MODEL.md

File metadata and controls

119 lines (80 loc) · 3.15 KB

Threat Model

Goal

This document gives a practical threat model for Code Summarizer so reviewers can quickly understand what the tool is trying to protect, what assumptions it makes, and where its limits are.

Assets to Protect

  • Sensitive source code snippets
  • Hardcoded secrets accidentally pasted into snippets
  • Internal implementation details
  • Developer workflow privacy
  • Trust in local-only analysis behavior

Primary Security Goals

  • Keep code analysis local to the user's machine
  • Reduce accidental exposure of secrets
  • Avoid dependency on internet-hosted AI services
  • Make data flow simple enough to review and audit

Threat Actors

  • Accidental user mistakes, such as pasting secrets
  • Unapproved external services that would receive code if the app sent data remotely
  • Malicious or risky dependencies
  • Misconfigured local environments
  • Unapproved or untrusted local model artifacts

Trust Assumptions

This project assumes:

  • The local machine itself is trusted enough to run the application
  • Ollama is installed from an acceptable source
  • The local model files are acceptable for the environment
  • The workstation is governed by local endpoint controls
  • The user understands that LLM output can be incorrect

Main Data Flow

  1. User pastes code into the desktop app
  2. App scans for likely secrets
  3. App optionally redacts detected secrets
  4. App builds a prompt locally
  5. App sends the prompt to the local Ollama instance on localhost
  6. App validates the returned JSON structure
  7. App renders the structured response locally

Threats Considered

Data Exfiltration to Cloud Services

Mitigation:

  • App design is local-first
  • Intended model communication is localhost only
  • No external AI API integration is built into the application

Residual risk:

  • Host or local runtime misconfiguration
  • Future code changes

Accidental Secret Exposure

Mitigation:

  • Secret scanning
  • Redaction support
  • Sensitive Mode can enforce redaction and hide raw model output

Residual risk:

  • Regex-based scanning is not exhaustive
  • User can still paste more than necessary

Unsafe Trust in Model Output

Mitigation:

  • Structured JSON schema validation
  • Clear documentation that validation is model-guided, not compiler-backed

Residual risk:

  • LLMs can hallucinate, miss issues, or be overconfident

Supply Chain Risk

Mitigation:

  • Open source code review path
  • Lockfiles present in the repo
  • Local build option

Residual risk:

  • Third-party dependencies still require organizational review
  • Local model runtime and model weights are separate trust decisions

Out of Scope

The application does not currently attempt to solve:

  • Formal classified-environment accreditation
  • Enterprise identity management
  • Compiler-accurate validation for every supported language
  • Centralized audit and SIEM integration
  • Guaranteed detection of every secret type

Recommended Hardening

  • Build from source in a controlled environment
  • Review dependencies before approval
  • Restrict to approved workstations
  • Approve Ollama and model files separately
  • Use Sensitive Mode by default for higher-risk work
  • Re-review changes before updating deployed versions