The Helios project is committed to delivering a secure, robust foundation model ecosystem. We treat vulnerability management with the highest priority and appreciate the responsible disclosure of any security concerns.
Security fixes and patches are applied to active minor releases. Older unsupported branches will not receive backported security patches.
| Version | Supported |
|---|---|
| 0.1.x | Yes |
| < 0.1.0 | No (Pre-release) |
Please do not report security vulnerabilities via public GitHub issues, pull requests, or public discussions.
If you discover a vulnerability in Helios, please report it via one of the following secure channels:
- GitHub Private Vulnerability Reporting: Submit an advisory directly via the Security tab of the Helios repository.
- Email: Send an encrypted report to
security@helios.ai.
To assist in rapid triage and remediation, please include:
- A descriptive summary of the vulnerability.
- Steps to reproduce or a minimal Proof-of-Concept (PoC) script.
- The component affected (
helios.model,helios.memory,helios.training,helios.tokenizer, CLI). - Assessment of potential impact and vector of exploitation.
The Helios core security team adheres to the following timelines:
- Initial Acknowledgment: Within 48 hours of report receipt.
- Triage & Preliminary Assessment: Within 7 calendar days, confirming exploitability and severity rating (CVSS v3.1).
- Remediation & Coordinated Disclosure: Security patches are developed, tested, and released within 90 calendar days, coordinated with the reporter prior to public disclosure.
- Model Weight Deserialization: Arbitrary code execution or unsafe memory access during model checkpoint or safetensors loading.
- Hierarchical Memory Architecture (HMA) Isolation: Cross-tenant data leaks, SQL injection in Episodic SQLite storage, or prompt-injection attacks traversing semantic retrieval buffers into execution contexts.
- Tool Execution & Procedural Sandboxing: Sandbox escapes, path traversals, or privilege escalation during procedural tool invocation.
- Training Controller Integrity: Tampering with atomic checkpointing (
.pt.tmp->.pt) or bypassing SHA-256 integrity checks during resumption.
- Attacks requiring physical access to the host or root-level compromise of the operating system.
- Volumetric Distributed Denial of Service (DDoS) against public demonstration endpoints.
- Vulnerabilities in third-party dependencies unless directly exploitable through Helios standard configurations.
- Theoretical attacks without practical proof-of-concept or actionable impact demonstration.
Helios supports responsible security research conducted in good faith. If you conduct vulnerability research in compliance with this policy:
- We will not pursue legal action against you or request law enforcement intervention.
- We will work collaboratively with you to understand and resolve the vulnerability.
- We will publicly credit your contribution in our security advisory (unless you request anonymity).