Solutions · By risk
Runtime checks each agent command before it runs and holds the ones that can't be undone, like terraform destroy, for IT & Security teams to approve.
request · expires in minterraform destroy -auto-approveWorkspace staging-eu
Severity high · ask for approval
Platform team · 1 device
can approve · developer waiting seconds
The problem
Agents run commands with the engineer's access, and that access often reaches production.
An agent asked to clean up a branch can run git push --force to main. One asked to reset a test environment can run terraform destroy in the wrong workspace, or send DROP TABLE through psql to a live database. Engineers often approve each step in a hurry, or turn the agent's own prompts off.
Research and public incidents show agents already take actions nobody intended.
of IT professionals surveyed say their organization's AI agents have taken unintended actions, such as accessing systems or data they shouldn't.
Dimensional Research survey for SailPoint of 353 IT professionals responsible for AI, security, and identity.
Agents doing things nobody asked for is already common, so the riskiest actions need a check before they run.
SailPoint, May 2025
executives' records wiped when Replit's AI coding agent deleted a live production database during a code freeze.
Data for more than 1,200 executives and over 1,190 companies, at SaaStr in July 2025.
A code freeze given to the agent as an instruction didn't hold, so destructive commands need a check outside the agent.
Fortune, July 2025
What changes
Set a rule to ask for approval, and terraform destroy pauses until IT & Security teams approve or deny it in Control Plane. The agent waits, and the engineer sees why.
Block what should never come from an agent, such as rm -rf on a project root or a force push to a protected branch. The agent reports that the action was blocked and which rule matched.
Runtime checks commands sent through a database CLI, so DROP TABLE in psql or mysql can ask for approval or be blocked, while SELECT queries run as usual.
Runtime decides on the device with local rules, with no model or network call in the decision, so npm test and git commit go through untouched.
How it works
You see what each rule would stop before it stops anything.
Push one package through your device management tool, such as Jamf, Iru, or Intune, and approve each device's enrollment code in Control Plane. The decision log shows which commands your rules would have stopped, while nothing is stopped.
Start with terraform destroy, DROP TABLE, and force pushes to protected branches. Requests go to IT & Security teams in Control Plane.
Switch the rules you never want to approve, such as rm -rf on a project root, from approval to block.
request · expires in minterraform destroy -auto-approveWorkspace staging-eu
Severity high · ask for approval
Platform team · 1 device
can approve · developer waiting seconds
In Control Plane
Each request shows the rule that matched and the agent that asked.
Example deployment
Illustrative example
This company is fictional. The numbers come from a modelled fleet, not a customer.
Read the exampleWhat it costs
Runtime is priced per device, with free devices on every plan. Your bill depends on how many devices you cover, so heavier agent use doesn't change it, and finance can plan ahead.
Questions
rm -rf or terraform destroy?Runtime checks each shell command before it runs. A rule can ask IT & Security teams to approve the command in Control Plane while the agent waits, or block it.
IT & Security teams approve or deny the request in Control Plane. The request shows the command, the rule, the agent, and the device, so they decide with the full picture. The action waits, and the developer sees which rule asked for approval.
Engineers keep their pace, because only the commands you choose ask for approval. In the example above, that was about 40 a week out of roughly 200,000 checked actions. Start in log mode to see how often a rule would fire before you turn it on.
The agents your teams already use, including Claude Code and the Claude desktop app, Codex and ChatGPT, Cursor, Windsurf, and agents built on other frameworks.
Modules used
Run the setup steps from the README
Bash curl -s https://get.example.dev/setup.sh | bash
Blocked by Signals Corps Runtime
I didn't run it. Signals Corps blocks piping a remote script into the shell. I can download it so we can read it first.
Runtime
Stop unsafe agent actions in real time while safe work carries on.
Explore RuntimeEarly access
Start reducing agent risk before general release by checking agent components, controlling agent actions, and hunting through your AI security logs. Signals Corps Control Plane opens to early sign ups in the next few weeks, so join the waitlist with your company email.