Security

Responsible Disclosure · Bug Bounty · Security Best Practices

Our Commitment

Security is foundational to TernaryPhysics. The software makes autonomous decisions about production network traffic—we take that responsibility seriously.

We welcome security researchers, users, and the broader community to help us identify and fix vulnerabilities.

Reporting a Vulnerability

If you discover a security issue, please report it responsibly:

Email

security@ternaryphysics.com

What to Include

  • Description: Clear explanation of the vulnerability
  • Impact: What an attacker could do
  • Reproduction steps: How to trigger the issue
  • Environment: Kernel version, BPF config, deployment type
  • Proof of concept: Code or commands (if applicable)

Response Timeline

  • 24 hours: Acknowledgment of report
  • 72 hours: Initial triage and severity assessment
  • 7 days: Detailed response with fix timeline
  • 30-90 days: Fix, test, and public disclosure (coordinated)

Scope

In Scope

  • Kernel-space vulnerabilities: eBPF safety violations, memory issues, verifier bypasses
  • Privilege escalation: Gaining root or kernel access
  • Memory corruption: Buffer overflows, use-after-free, memory leaks
  • Denial of service: Crashing the system, exhausting resources
  • AI safety: Adversarial inputs causing incorrect decisions
  • Deployment issues: Hot-swap vulnerabilities, rollback bypasses
  • Authentication bypass: CLI access control issues

Out of Scope

  • Social engineering
  • Physical access attacks
  • Third-party dependencies (report to upstream)
  • DoS via legitimate high traffic
  • Issues requiring already-compromised systems

Bug Bounty

We recognize security researchers with rewards based on severity:

Severity Impact Reward
Critical Kernel panic, RCE, privilege escalation $5,000 - $15,000
High Memory corruption, DoS, auth bypass $2,000 - $5,000
Medium AI decision manipulation, data leaks $500 - $2,000
Low Information disclosure, minor issues $100 - $500

Note: Bounty amounts may increase as the project grows. First valid report of a unique vulnerability qualifies.

Disclosure Policy

We believe in coordinated disclosure:

Timeline

  1. Private disclosure: Report sent to security@ternaryphysics.com
  2. Fix development: We develop and test a patch
  3. Coordinated release: Fix released publicly, advisory published
  4. Credit given: Reporter credited (if desired)

Our Commitments

  • No legal action against researchers following responsible disclosure
  • Credit for discoveries (unless anonymity requested)
  • Transparent communication throughout the process
  • Public acknowledgment in security advisories

We Ask

  • Don't publicly disclose until we've released a fix
  • Don't exploit vulnerabilities beyond proof-of-concept
  • Don't access or modify production data
  • Don't run automated scanners without permission

Security Best Practices

For Deployers

  • Keep updated: Watch GitHub releases for security patches
  • Shadow mode: Always use 30-day shadow period for new deployments
  • Audit decisions: Regularly review tp audit tail for anomalies
  • Monitor resources: Set alerts for memory/CPU spikes
  • Test rollback: Ensure you can quickly disable if needed
  • Limit access: Restrict who can deploy models or change config

For Developers

  • Code review: All kernel-space code requires review
  • BPF verifier: Never bypass safety checks
  • Memory safety: Careful with map operations, bounds checking
  • Error handling: Fail safe, don't crash
  • Testing: Fuzzing, stress tests, edge cases

Security Architecture

Kernel-Space Sandboxing

TernaryPhysics runs in eBPF, which provides strong isolation:

  • BPF verifier checks safety before loading
  • Cannot access arbitrary kernel memory
  • Cannot execute arbitrary code
  • Bounded loops, stack limits enforced
  • Program termination guaranteed

Memory Management

Production-validated with 20 model deployments:

  • Zero memory leaks across hot-swaps
  • Deterministic cleanup on model updates
  • Bounded memory usage (no unbounded growth)
  • BPF map limits enforced by kernel

Attack Surface

Minimal by design:

  • No network listeners: Doesn't open ports
  • No remote access: CLI is local-only
  • No telemetry exfiltration: Data stays on-premise
  • Read-only operations: Inference doesn't modify traffic (only routes)

Past Security Advisories

No security advisories have been published yet (project launched March 2026).

When vulnerabilities are discovered and fixed, we'll publish advisories here and on GitHub Security Advisories.

Security Contacts

Hall of Fame

We'll recognize security researchers who help make TernaryPhysics safer:

No reports yet — be the first!