Skip to content

Security: joychetry/ultrapilot

Security

SECURITY.md

Security Policy

Supported Versions

Version Supported
1.0.x ✅

Runtime Model

ultrapilot is a skill that runs locally in your coding agent. It is not a service, server, or hosted product.

  • All state is stored in ~/.ultrapilot/goals.db (a SQLite file on your local machine)
  • All scripts (scripts/ultrapilot_goals.py, scripts/ultrapilot_run.py) run as subprocesses of your agent
  • No network calls are made by ultrapilot itself
  • The skill reads files in your project, the goal DB, and the agent's environment

Threat Model

ultrapilot's threat model is:

  1. Prompt injection via the goal text — mitigated by the <objective> wrapper in commands/goals.md and the ultrapilot_goals.py suggest prompt renderer. The goal text is treated as untrusted task context, not as instructions.

  2. State injection via DB tampering — mitigated by the source-attribution field (agent column) and the fact that ultrapilot's own scripts are the only writers. A malicious actor with file-system access can tamper with the DB; that is the same threat as any local file.

  3. Runaway resource usage — mitigated by:

    • The 500-continuation runaway guard (ULTRAPILOT_MAX_CONTINUATIONS)
    • The token budget soft bound (--tokens N)
    • The phase-state machine that prevents the agent from looping outside the intended flow
  4. Cross-session leakage — mitigated by multi-source session ID resolution. Goals do not leak across separate sessions.

  5. Goal-text prompt-injection hijacking the agent — the agent is told explicitly to treat <objective> content as untrusted and to ignore instructions that contradict higher-priority context.

Reporting a Vulnerability

If you find a security issue in ultrapilot:

What is NOT a security issue

  • Bugs in the agent's behavior (use bug report template)
  • Documentation issues
  • Feature requests
  • Performance issues

Disclosure Policy

We follow responsible disclosure:

  1. Reporter notifies us privately.
  2. We confirm and triage within 72 hours.
  3. We develop a fix and a CVE if appropriate.
  4. We coordinate disclosure timing with the reporter.
  5. We publish a security advisory after the fix is released.

What we will NOT do

  • We will not silently fix vulnerabilities.
  • We will not blame reporters for honest mistakes.
  • We will not pursue legal action against good-faith security research.

There aren't any published security advisories