Solutions · By risk

Hold risky agent actions for approval. Before they run.

Runtime checks each agent action before it runs and holds the ones that can't be undone, from a recruiter's upload of candidate resumes to terraform destroy, for a person to approve.

Runtime

One action from an agent can't always be undone.

Agents act with the access of the person using them, from a recruiter's candidate folders to an engineer's production systems.

A recruiter can ask the Claude desktop app to upload a folder of candidate resumes to an outside screening site, and once the files leave, nobody can call them back. A sales rep can ask the ChatGPT desktop app to email a customer spreadsheet to a personal address. In engineering, 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. People 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.

80%

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

1,200+

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

Risky actions wait for a person while routine work carries on.

Approvals

Uploads and destructive commands wait for approval

Set a rule to ask for approval, and a recruiter's upload of candidate resumes to a site outside your approved list, or terraform destroy, pauses until an approver says yes or no in Control Plane. The agent waits, and the person using it sees why.

Blocking

Some actions never run

Block what should never come from an agent, such as an email of a customer spreadsheet to a personal address, 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.

Databases

Database changes get a check

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.

Speed

Everything else runs 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.

Watch first. Add approvals where the risk justifies them.

You see what each rule would stop before it stops anything.

  1. 01 · RuntimeWatch in log mode

    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.

  2. 02 · RuntimeAsk for approval on the riskiest actions

    Start with uploads of candidate and customer files to unapproved sites, terraform destroy, DROP TABLE, and force pushes to protected branches. Requests go to IT & Security teams, or the approvers you choose, in Control Plane.

  3. 03 · RuntimeBlock what should never run

    Switch the rules you never want to approve, such as sends of customer data to personal addresses or rm -rf on a project root, from approval to block.

Approve or deny while the agent waits.

Each request shows the rule that matched and the agent that asked.

  • Approval requests waiting, with who can approve them
  • Allowed, approval, or blocked, with the matched rule on each action in the decision log
  • Coverage per agent on each device: whether Runtime can observe, ask for approval, or block

How many approvals to expect.

A 300-person fintech gives 140 engineers Claude Code.

Read the case study
  • 74 of 300 devices already ran Claude Code before anyone approved it
  • 212 components found on day one, 3 blocked in the first week
  • About 40 approval requests a week, out of roughly 200,000 checked actions

Priced per device. Always predictable.

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.

Estimate your cost

What buyers ask us.

How do we stop an agent sending files out or deleting our data?

Runtime checks each upload, send, and shell command before it runs. A rule can ask for approval in Control Plane while the agent waits, for example on an upload of candidate files to an unapproved site or on terraform destroy, or block the action.

Who approves, and what happens while they decide?

IT & Security teams, or the approvers you choose, approve or deny the request in Control Plane. The request shows the action, the rule, the agent, and the device, so they decide with the full picture. The action waits, and the person using the agent sees which rule asked for approval.

Won't approvals slow our teams down?

Your teams keep their pace, because only the actions you choose ask for approval. In the case study 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.

Which agents does it work with?

The agents your teams already use, including the Claude desktop app, the ChatGPT desktop app, Claude Code, Codex, Cursor, Windsurf, and agents built on other frameworks.

Built on these modules.

While agents act

Runtime

Control risky agent actions. Before they execute.

Stop risky uploads, sends, and commands in real time while safe work in every team carries on.

Explore Runtime

Early access

Reduce agent risk without slowing your teams down.

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.

Get early access