For developers using AI coding agents
Prompt injection can turn your AI agent against you. SigmaShake is the last line of defense.
A poisoned README, web page, or tool result can hijack your coding agent. Suddenly it’s deleting files, reading your secrets, or sending data to an attacker. You can’t reliably catch every injection. But you can stop what it tries to do.
You can’t read the attacker’s mind, so guard the action
Text-level filters try to spot the malicious prompt, and they miss things, because attackers are endlessly creative. SigmaShake takes the opposite, last-line approach: it doesn’t try to detect the injection. It inspects the action the agent is about to take and refuses the dangerous ones, deterministically, before they run. A matching rule can still stop an injected action from deleting your repo or reading your keys. A call that no rule matches is allowed, and ssg is not a sandbox.
What it gates
- Destructive commands (recursive deletes, database drops)
- Secret reads and outbound exfiltration (Egress Guard)
- Tampering with your agent’s own config and credentials
SigmaShake runs locally and deterministically, with the same behavior across Claude Code, Cursor, and Codex plus decisions recorded to a local audit log on a best-effort basis.
30 seconds to protected
1. Install
curl -fsSL install.sigmashake.com | sh -s -- -y2. Wire it in
ssg initRunning -y accepts the Terms of Service. An action gate is a complement to input/output filtering, not a replacement; run both for defense in depth. See the docs.