HiveGuard Security Core is a framework for building secure autonomous AI agents. Security is the product. If you find a weakness, we want to hear from you.

Please do not open a public GitHub issue for security problems.

Email security@ubava.ee (or info@ubava.ee if the security address is unavailable) with:

• A clear description of the issue
• The affected module(s) and version
• Reproduction steps or a proof-of-concept if possible
• Your assessment of impact (confidentiality / integrity / availability)
• Whether you want public credit after the fix lands

We will acknowledge receipt within 72 hours. If you do not hear back in that window, follow up — do not assume the report was received.

Anything in this repository — the public hiveguard-security-core framework:

• framework/agents/base_agent.py — base class, health server, graceful shutdown, memory wipe
• framework/lifecycle/states.py — state machine + transition table
• framework/security/invariants.py — the 20-invariant spec
• framework/security/enforcement.py — runtime enforcement decorators (@enforces, sanitize_input, validated_urls, etc.)
• framework/types/ — type definitions
• examples/echo_agent.py, examples/monitor_agent.py

Examples of in-scope findings:

• Framework invariants that can be bypassed without raising InvariantViolation
• Decorator stacks that silently drop violations
• safe_url() false negatives (URLs that should be blocked but aren't)
• sanitize_input() depth or escape-sequence bypasses
• Memory leaks that survive stop() with @memory_wiped_on_exit
• Any exploit that lets an attacker bypass the 20-invariant contract

• The commercial HiveGuard agents (Scout, Analyst, Fixer, WebScout) available via ubava.ee — these have their own private disclosure channel. Contact security@ubava.ee and we will route the report to the right team.
• The UBava Relay — reports about the privacy relay, VHH air-lock, PII cascade, or Docker cascade infrastructure should also go to security@ubava.ee.
• Dependencies — we have zero runtime dependencies. Dev-only dependencies (pytest) should be reported upstream to the respective project.
• Attacks that require local machine compromise already (root on the machine running the framework).
• Social engineering or physical attacks.

After acknowledging your report, we aim to:

We practice coordinated disclosure: we ask reporters not to publish details until a fix is live, and in exchange we commit to a concrete timeline.

This is an early-stage framework. During the 0.x series, only the latest minor release receives security fixes.

When 1.0.0 ships, the previous major version will continue to receive critical security fixes for six months.

Reporters who follow responsible disclosure will be credited in the commit that fixes the issue and in the release notes, unless they request anonymity.

HiveGuard Security Core is the open-source spec and runtime enforcement layer. The 20 invariants in framework/security/invariants.py describe the security contract every HiveGuard agent — open or commercial — must satisfy.

The commercial agents behind the UBava Relay enforce the full spec end-to-end, including the 8 invariants that can't be verified in a pure-Python framework skeleton (jurisdiction-aware autonomy, inter-agent authentication, BDP integrity, etc.). If your research touches those commercial components, the same security@ubava.ee address reaches the team that owns them.

Thank you for helping us keep autonomous agents safe. 🐝