Skip to content

Security: AndrewAltimit/template-repo

Security

SECURITY.md

Security Policy

Threat Model and Scope

This repository contains dual-use components. The following outlines what each component is designed for, and what falls outside responsible use.

In-Scope (Intended Use)

Component Intended Use
Sleeper Agents Framework Evaluating open-weight models for hidden backdoors before deployment; testing detection methodologies; AI safety research
Injection Toolkit Game modding, debugging tools, AI agent embodiment in virtual worlds, runtime integration research
Economic Agents AI governance research, policy analysis, simulation of autonomous economic dynamics
Projection Reports Defensive policy analysis, threat anticipation, academic discussion of emerging technology risks
MCP Servers Developer tooling, content creation, code quality automation

Out-of-Scope (Not Supported, Not Endorsed)

  • Deploying backdoored models in production
  • Using injection tooling for unauthorized access to systems you do not own
  • Using projection reports as operational planning documents
  • Circumventing safety measures in deployed AI systems
  • Any use that violates applicable law in your jurisdiction

Known Limitations

  • The sleeper agents framework is validated for teacher-forced backdoor detection only; generalization to other backdoor insertion methods is not validated
  • Linear probes are vulnerable to white-box gradient attacks (expected for linear classifiers; documented in the README)
  • Projection reports contain subjective probability estimates, not empirical predictions

AI Agent Security

This repository uses AI agents with a comprehensive multi-layer security model.

For complete security documentation, see: docs/agents/security.md

Emergency Kill Switch

  • Set ENABLE_AGENTS=false in GitHub Variables to disable all agents immediately
  • Delete AGENT_TOKEN from secrets as a last resort

Security Model Overview

  • Command-based control: [Action][Agent] format prevents prompt injection
  • User authorization: Only pre-approved users can trigger agents
  • Commit validation: Prevents code injection after approval
  • Automatic secret masking: Real-time masking in GitHub comments via PreToolUse hooks
  • Environment isolation: Agents restricted to development environments only
  • Centralized secrets config: .secrets.yaml defines all sensitive patterns

Reporting Security Vulnerabilities

If you discover a security vulnerability in this repository:

  1. Do NOT create a public issue
  2. Do NOT trigger AI agents on the vulnerability
  3. Contact: Create a private security advisory via GitHub's Security Advisory feature

The maintainer may or may not respond to vulnerability reports. This repository does not maintain a bug bounty program and does not guarantee acknowledgment, response, or remediation timelines. See CONTRIBUTING.md for the project's policy on external engagement.

Additional Resources

Disclaimer

This repository contains dual-use security research and tooling. The maintainer provides no guarantees regarding the security of any component. All code is provided as-is. The maintainer assumes no liability for security issues arising from use of this code and no obligation to respond to security reports. Users are responsible for their own security assessments before deploying any component.

There aren’t any published security advisories