DevOps Copilot connects to real infrastructure over SSH, stores encrypted server credentials, and can execute commands on remote hosts. We take vulnerabilities in this project seriously and appreciate responsible disclosure.

Only the latest published release on PyPI receives security fixes. We do not currently maintain long-term-support branches.

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

Instead, report it privately using one of the following:

• GitHub Security Advisories (preferred)
• Email the maintainer (see the GitHub profile linked in the repo) with the subject [SECURITY] devops-copilot

Please include:

• A description of the vulnerability and its impact
• Steps to reproduce (a minimal PoC helps a lot)
• The affected version/commit
• Any suggested remediation, if you have one

We aim to acknowledge reports within 72 hours and to provide a fix or mitigation plan within 14 days for confirmed issues, depending on severity.

Given what this project does, the following areas carry higher risk and get extra scrutiny:

• Credential handling — SSH passwords/keys are encrypted at rest with Fernet (AES-256) in app/core/security.py. Report any issue that could leak plaintext credentials, weak key derivation, or key reuse.
• Command guardrails — app/modules/guardrails/service.py uses semantic similarity search to block dangerous commands before execution. Bypasses of this guardrail (crafted commands that should be blocked but aren't) are considered security issues, not just bugs.
• Human-in-the-loop approval — Write commands require explicit [y/N] approval via LangGraph's interrupt. Any path that executes a write command without approval is a critical issue.
• Auth/JWT — app/modules/auth/ issues and validates JWTs. Report issues with token forgery, expiry bypass, or privilege escalation.
• Prompt injection — Since an LLM agent has tool access to SSH execution, report any way that content returned from a remote server (logs, file contents, command output) can manipulate the agent into executing unintended commands, bypassing guardrails, or exfiltrating credentials.

We follow coordinated disclosure: once a fix is available, we'll credit reporters (unless they prefer to stay anonymous) in the release notes and CHANGELOG.md.