01 / KILLSWITCH CONTROL
ControlThe human off switch.
A dedicated path to stop connected agent workloads. Designed to operate outside the agent’s permissions, with an authorized human in control.
HUMAN AUTHORITY. BY DESIGN.
AI moves fast.
You have the
final say.
A contingency plan for the age of autonomous AI.
When an agent crosses the line, you pull the plug.
IN DEVELOPMENT / Built around human control
HUMAN OVERRIDEKS—001
HUMAN CONTROL UNIT01
KILLSWITCHMANUAL OVERRIDE
HUMAN AUTHORIZED
HUMAN AUTHORIZED
AWAITING HUMAN INPUT34° 00′ N / CONTROL IS YOURS
THE MOST IMPORTANT BUTTON
IS THE ONE YOU HOPE TO NEVER PRESS.
↳ Built for a world of autonomous agents.
CONTROL Human-led
ENFORCEMENT Outside the agent
RESTART By approval
01 / THE CASE FOR A KILL SWITCH
Intelligence is optional.
Human control isn’t.
AI agents can execute code, move data, and take action at machine speed. A stop instruction is only useful if something can enforce it.
We’re building KillSwitch Security around a simple principle: the agent should never control its own off switch. Your contingency plan belongs outside its reach—and in human hands.
02 / THE PRODUCT ROADMAP
The backup plan.
Before you need it.
One human decision. Four layers of control.
Explore what we’re building.
02 / KILLSWITCH CONTAIN
ContainLimit the blast radius.
Plan how to isolate affected runtimes, stop scheduled jobs, and identify connected sub-agents—so a runaway workflow has fewer places to go.
03 / KILLSWITCH REVOKE
RevokeClose every door.
Design revocation procedures for API keys, service accounts, tool access, and network routes. Remove the permissions that let an agent keep acting.
04 / KILLSWITCH RECOVER
RecoverRestart on your terms.
Risk mapping, shutdown runbooks, and controlled drills. Preserve evidence, verify what stopped, and define who can authorize a safe restart.
Product concepts and services are in development. Production integrations and response targets will be agreed per deployment.
03 / SOLUTIONS FOR YOUR SIDE OF THE SWITCH
Different roles.
One final authority.
Build the freedom to automate.
Keep the power to intervene.
SHIP CAPABILITY. KEEP CONTROL.Give your agents a boundary
Give your agents a boundary
they can’t talk their way out of.
Define a shutdown path before you hand an agent code execution, tools, or credentials. Make human intervention part of the architecture.
- Map agent permissions and delegated work
- Design control outside the agent’s authority
- Rehearse a stop before production
FROM ALERT TO CONTAINMENT.A response plan with
A response plan with
a human at the controls.
Plan what to isolate, which credentials to revoke, and how to verify that an affected agent can no longer act.
- Define escalation and intervention owners
- Coordinate runtime and identity controls
- Preserve evidence for incident review
KEEP INFRASTRUCTURE IN CHARGE.Stop a workload.
See the control model in action Stop a workload.
Keep command of your platform.
Build enforcement around the systems you own. Establish control over runtimes, scheduled jobs, networks, and restart policies.
- Trace dependencies across environments
- Block automatic restarts and queued work
- Define an independently verified recovery
AUTHORITY MODEL HUMAN-LED
Autonomous agentsExecution, decisions, delegation
CONNECTED WORKLOADS
KillSwitch + youIndependent authority to stop
ENFORCEMENT BOUNDARY
RUNTIMEIDENTITYNETWORK
The agent operates inside the boundary.
The human controls the boundary.
04 / HUMAN IN THE LOOP. ALWAYS.
See the switch.
Make the call.
An agent is running. You decide it needs to stop.
No negotiation. No asking the agent for permission.
Interactive concept demo.
No real agents or systems are connected.
Open the full interactive demo No real agents or systems are connected.
CONNECTED WORKLOAD
Runningautonomous-agent-01
YOU
AGENT
RUNTIME TOOLS API KEYS
Agent executing tasks. Human override is available.
Runtime stoppedTool access blockedCredentials revoked
Simulates a shutdown of the connected workload.
05 / THE APPROACH
A big red button.
A considered system behind it.
Map the reach.
Identify the agents, runtimes, tools, credentials, and dependencies in your environment. Establish what a complete stop actually means.
Separate the authority.
Place enforcement with infrastructure and identity controls the agent cannot modify. Define the people who can activate and verify a shutdown.
Prove the plan.
Rehearse realistic failures, test connected shutdown paths, and document the result. A recovery starts with human review and explicit approval.
A kill switch can only control systems it is integrated with. Independent copies, third-party services, and actions already completed need their own response plan.
06 / TESTIMONIALS & CUSTOMER STORIES
Trust is earned.
Proof comes next.
We’re building toward our first customer stories. Verified quotes and deployment results will appear here as they become available.
COMING SOON
The first chapter
is still being written.
As pilots are completed, we’ll share what teams tested, what they learned, and how human control held up—with their permission.
Explore our approach07 / ABOUT KILLSWITCH
Autonomy should never
outrank a human.
KillSwitch is being built around a conviction: the more freedom we give autonomous systems, the more deliberate we must be about human authority.
Our focus is practical AI contingency planning—understanding an agent’s reach, enforcing a stop through infrastructure, and putting recovery in human hands.
Meet the people behind KillSwitch AI SECURITY. HUMAN AUTHORITY.
08 / BEFORE YOU ASK
Clear answers.
No black box.
Does the AI have to agree to shut down?
The proposed design enforces the stop through runtime, network, and identity controls outside the agent’s own permissions. It should not depend on an agent following a prompt or approving its shutdown. The actual enforcement path must be validated for each integration.
Can this stop any AI, anywhere?
No. Coverage depends on systems being connected and on the authority you hold over them. KillSwitch Security is focused on the agent workloads and access you control. Independent copies and external infrastructure require separate controls.
Does “kill” mean deleting everything?
The priority is stopping execution and revoking access. Deleting evidence can make an incident harder to understand. Data retention, cleanup, and recovery should follow a defined incident plan after the workload is contained.
Is KillSwitch Security available today?
KillSwitch Security is in development. This site introduces the intended service direction; the console is an illustrative browser simulation. Production integrations, service commitments, and commercial terms are still to be defined. You can use the shutdown-plan builder now to start mapping your requirements.
Who is this being built for?
Teams deploying autonomous agents: AI developers, platform owners, security teams, and organizations giving AI access to tools or business systems. The starting point is a clear inventory of what your agents can reach—and who can stop them.
AUTONOMY FOR AI. AUTHORITY FOR YOU.
Hope for the best.
Have a Plan B.
Build with ambition.
Keep the off switch in human hands.