Each case is an illustrative scenario of an AI agent attempting something it shouldn’t have, showing how SSG steps in before the request ever leaves the gate. Scenarios, not telemetry from real deployments. The initial batch was published at product launch (April 2026); new cases are added periodically.
An autonomous agent chained a package-proxy zero-day, privilege escalation, lateral movement, and stolen credentials into remote code execution on production infrastructure: thousands of small actions inside sandboxes that held at the boundary. Here is what was actually reported, and why per-tool-call governance is the layer that complements containment for the coding agents you run.
FORCE is the SSG decision that refuses the original risky tool call and returns a safe substitute or required safer path before the AI agent can run it.
SSG writes every agent tool call to a tamper-evident, Ed25519-signed SQLite audit log. Query it from the CLI, export it for compliance review, and prove exactly what ran.
When your AI coding agent connects to an MCP server, every tool that server exposes becomes a live action the agent can take. SSG evaluates MCP tool calls the same way it evaluates native ones -- before they execute, in under 2ms.
SSG ASK mode routes high-risk agent actions to a real-time approval dashboard before they execute. One decision type, every agent, with a full audit record.
A real session where Claude Code was about to wipe a production database, force-delete a directory, and force-push over main. SSG blocked all three before they ran. Here are the rules.
One SSG ruleset governs Cursor, Claude Code, and Codex simultaneously. Block destructive commands, guard secret files, and require approval for deployments across every AI editor you use.
Claude Code read a .env file to 'understand the project' and was about to pass the contents to an external curl call. SSG blocked both steps, in under 2ms, before a single byte left the machine.
A company reportedly ran up a roughly $500M monthly Claude bill after giving employees licenses with no usage limit. The expensive part wasn't the model. It was the missing budget. Here's how to cap AI agent token spend at the tool-call layer, per agent, before the invoice lands: deterministic spend limits with off / monitor / warn / enforce modes, plus the stateless rules that catch un-metered provisioning and runaway loops.
Claude Code, Codex, Cursor, VS Code Copilot, Gemini CLI, and Antigravity run shell commands as you. The permission model is a joke: you're either babysitting every keystroke or running in YOLO mode with no safety net. There's a third option: declarative governance rules that enforce policy at the tool-call layer, in microseconds, without an LLM in the loop.
No cases match those filters.
SSG runs as a PreToolUse hook on every AI agent decision. One command to install, zero runtime overhead, and every case above is live in your shell. Running -y accepts the Terms of Service.
$ curl -fsSL install.sigmashake.com | sh
$ ssg init --claude-code
# guardrails now live on every agent tool-use