Alphacode is an AI coding agent that runs arbitrary code, including shell commands proposed by a language model. While the agent has built-in safety gates for destructive commands (see bash_destructive_gate.rs), it is still a powerful tool. Please use it responsibly and report vulnerabilities responsibly.

Only the latest release receives security fixes. Please upgrade before filing a report.

Please do not file a public GitHub issue for security bugs.

Email: tlbbeg313@gmail.com with the subject prefix [SECURITY].

Include:

1. A clear description of the vulnerability and its impact.
2. Reproduction steps (a transcript, a PoC repo, or a payload).
3. Your alphacode --version and platform info.
4. Whether you intend public disclosure and on what timeline.

You should receive an acknowledgement within 72 hours. We aim to ship a fix within 14 days for high-severity issues.

We are interested in real vulnerabilities. Please do not report:

• The AI model itself producing bad code (use your provider's safety channels).
• Hypothetical jailbreaks against the underlying model.
• "The agent could be tricked into running X" without a concrete repro.

Alphacode ships with several layers designed to limit blast radius:

• Destructive-command gate (bash_destructive_gate.rs): refuses catastrophic targets (/, $HOME, credential stores, device nodes) and asks for human justification on ambiguous commands.
• Risk assessment (alphacode_command_risk): tokenizes shell commands to classify risk before they execute.
• Permission prompts: the TUI surfaces every risky action for confirmation.
• Network URL safety: outbound HTTP fetches are classified against SSRF and credential-leak heuristics.
• Session crash recovery: a panic or signal labels the session as Crashed and prints a resume command so the user never loses work.

If you find a way to bypass any of these, please report it.