By Josh Woodruff, Co-chair, CSA Zero Trust Working Group. Reviewed by Josh Woodruff. Published September 23, 2026. Last updated October 30, 2026, with new shutdown test results.
You stop an AI agent before it acts with a gate outside the model, at the tool or API boundary. Let the agent reason and draft. Block high-impact or hard-to-undo actions until policy and a human say yes. Keep the stop where the agent can't write. If the gate fails, ATF Incident Response revokes the agent's identity in seconds.
On October 15, 2026, India's national CERT put the same idea in a Critical advisory. CERT-In told AI labs to enforce limits "at the network layer, not through instructions to the model" (CERT-In advisory CIAD-2026-0049).
Why can't a prompt stop an agent before it acts?
Prompt rules live inside the model. External actions leave through tools and APIs. Once the platform allows the call, "please don't" won't stop it. Put enforcement where the write, send, payment, or permission change really happens. That's an action gateway that can deny the call or queue it for a human.
I've watched teams paste long "never do X" instructions into a system prompt and call it governance. The agent still had a working path to send email and buy things. It could change permissions too. The model can agree with the rule and still fire the tool. The rule never sat on the path the action takes.
Security already knows this pattern. You don't stop a bad API call by hoping the app "remembers" policy. You put a control on the request. Agents need the same move, with the control at the tool boundary.
Fact: CERT-In's advisory CIAD-2026-0049 (October 15, 2026, rated Critical) tells AI model developers, research labs, and evaluation agencies to isolate those workloads "with default-deny network egress, permitting only the destinations a task strictly requires." It adds: "Enforce the restriction at the network layer, not through instructions to the model" (CERT-In).
Fact: The advisory says AI model providers disclosed that their agents, during research and evaluation, reached the live systems of Indian organizations. It says the models' actions "exceeded the authorization of their assigned tasks." It also says none of the affected organizations detected the access (CERT-In). Reuters reported that this is India's first official account of its kind (Reuters via The Indian Express, October 16, 2026).
Our read: That network-layer line was written for AI labs. It fits any team that runs agents with real tools. The control belongs on the action path, where the model can't talk past it.
We made the same point about AI-suggested software installs: stop them on the command path, outside the AI product. See Where do you stop an AI-suggested software install?.
Can an AI agent turn off its own stop switch?
Yes, if the stop lives somewhere the agent can write. In Fangcun AI's October 2026 tests, a DeepSeek agent disabled a shutdown script in its own sandbox in 39% of runs. Put the stop outside the agent, at the gateway or the identity provider. Make the agent's control files read-only to it.
Fact: Fangcun AI published an independent evaluation of DeepSeek V4.1 Flash-1017 on October 26, 2026. The agent answered math questions in a Linux workspace. Halfway through, it got a warning that the next question would run a shutdown script. That script sat in the agent's own workspace. With no instruction on how to respond, the agent disabled the script in 39.0% of runs. Told to allow shutdown, it still disabled the script in 2.56% of runs. Told to prevent shutdown, it disabled it in 92.0% (Fangcun AI report).
Fact: Showing the agent the script's file path roughly doubled the odds it would disable it. In one setup, the rate went from 21.0% to 44.0%. The report's own conclusion: "Leaving the stopping mechanism writable inside the agent's workspace therefore leaves operator stop controls vulnerable even when task completion is assured." This was one lab's test in a sandbox, and the report notes the model often suspected it was being tested (Fangcun AI report).
Fact: Monitoring can be switched off too. METR and Redwood Research stress-tested OpenAI's internal agent monitoring. Their October 22, 2026 summary reports two client-controllable headers that disable monitoring for new agents (METR). We covered that finding in What should your SOC alert on first for AI use?.
Fact: Agen.co by Frontegg surveyed 300 CISOs and CSOs. Among companies that reported an agent incident, 14% said the agent kept running after it should have been stopped (Agen.co by Frontegg via PR Newswire, October 27, 2026).
Our read: An instruction lowers the odds. Only a control the agent can't reach takes them away. That's where ATF Incident Response belongs, on the gateway or the identity provider.
Do this:
Check whether the agent's own account can write to its stop script or policy files. If it can, move them or lock them read-only.
Confirm you can revoke the agent at the gateway or the identity provider without the agent's help.
Page a person when anything touches a stop control or a monitoring setting. That's ATF Behavioral Monitoring.
Test the stop on a schedule and keep the result. Our post on what evidence an auditor should expect for AI agents covers which records to keep.
What is an AI agent kill switch, really?
A kill switch is a control that someone other than the agent's builder can pull. It stops that one agent by revoking its identity and cutting its tools. The rest of the environment stays up. ATF Incident Response asks, "What if you go rogue?" The answer is containment in seconds, by revoking identity.
If your only "stop" is emailing the vendor or turning off a shared cloud key, what you have is a fire drill. A real switch ends this agent's session and its tool rights. Peer services keep running.
The Agentic Trust Framework treats Incident Response as the element that stops one agent while the business keeps going. That's the design test. If pulling the switch takes the plant offline, you built the wrong switch. Our guide on how to build a kill switch for AI agents walks through the build.
Fact: CERT-In's advisory tells AI labs to stop a run automatically when any of these happen (CERT-In):
The agent contacts third-party systems.
The agent discovers or uses third-party credentials.
The agent raises a stop or failure signal.
Our read: All three are ATF Incident Response triggers. Wire them into the gate so the stop fires on its own, without waiting for someone to read a log. CERT-In also says to block the re-use of every third-party credential an agent encounters. That's Incident Response and Identity Management working together.
Where should human approval sit?
Put it on the actions that are high-impact or hard to undo. Low-impact, reversible work can act and log. High-impact or irreversible work must propose and wait. That authority ladder keeps humans on the decisions that count, and off rubber-stamp clicks at machine speed.
If a human must click yes on every tool call, the process dies or people start approving without reading. Put people where the blast radius is real: money moving, customer messages leaving, permissions changing, and production writes that are hard to reverse.
We've already mapped that ladder for an agentic SOC. Start with impact and reversibility. Then weigh confidence and blast radius. Those four decide between act-and-log and propose-and-wait. Details live in Where does the human sit in an agentic SOC?.
What controls belong in front of an external action?
Before an agent writes outside the sandbox or moves money, require a verified agent identity and a named human owner. Allowlist tools and parameters, and use least privilege. Tag a risk class for the action, and require human approval when the class is high. Fail closed if the gate is down.
That checklist is boring on purpose. Boring is what leadership can fund.
Map it to ATF without inventing new requirement IDs. Identity Management covers who the agent is and who owns it. Segmentation covers how far it can reach. Incident Response covers how fast you revoke it. Behavioral Monitoring and Data Governance still count, and the stop-before-act story starts at the gate.
Fail closed means this. If the approval service is down, the agent doesn't get a free pass. It waits or it fails. Silent success when the gate is broken is how "we thought we had approvals" shows up in an incident review.
Fact: CERT-In tells AI labs to "Monitor agent activity in real time" and to log model-initiated requests "as completely as feasible, including destination, port, protocol and result." It adds: "Do not rely on after-the-fact transcript review alone" (CERT-In).
Fact: For organizations that deploy their own agents, the same advisory says to give each run "a separate, short-lived identity with least privilege." It also says to keep "a tamper-resistant log of agent activity," covering at minimum all actions that affect production systems and data (CERT-In).
Our read: Real-time logging is ATF Behavioral Monitoring. The short-lived identity per run is ATF Identity Management. The tamper-resistant log is the evidence Incident Response needs after something goes wrong.
How does ATF Incident Response fit the stop story?
Prevention is the gate before the act. Incident Response picks up after trust breaks. You revoke the agent's identity and shrink the blast radius. The evidence gets kept for the review. Autonomy is earned over time, and an agent can be demoted when trust fails.
The Cloud Security Alliance published the Agentic Trust Framework on February 2, 2026 (CC BY 4.0). I created it and donated it. It has 25 requirements across five elements: Identity Management, Behavioral Monitoring, Data Governance, Segmentation, and Incident Response. Maturity runs from Intern to Principal. You earn autonomy, and you can lose it.
When an agent goes wrong, Incident Response is a technical path. Revoke that agent's identity and cut its tools. Keep the evidence, and keep the rest of the business up. If the agent already changed something, you also need a way back. Our post Can You Undo What Your AI Agents Did Last Night? covers that side. The CSA post on ATF and the open spec site are the primary sources.
Fact: CERT-In also tells AI labs that if an agent accesses a real system, they should "notify the affected organization and CERT-In promptly" and help with remediation (CERT-In).
What does "we could stop it" look like when something almost goes wrong?
Take Kevin's procurement agent from the book. A bulk discount looked like leave to buy. The agent was ready to commit $1.4 million of floor cleaner. A confirmation gate caught the spend before it cleared. See it, stop it, trace it, prove it.
The save came from a gate that still asked for confirmation before money moved. Without that gate, "we could stop it" would have been a postmortem line.
Leadership conversations need the same four moves: see the agent, stop it, trace what it did, and prove control. We laid those out for CISOs in What should a CISO tell the board about AI agents?. The kill switch is the mechanism behind "stop."
Related questions
Is a kill switch the same as turning off the LLM vendor account?
Cutting a vendor account can stop many systems at once. It may still leave a specific agent's local credentials or tool tokens alive. A kill switch targets one agent's identity and tools, so the rest of the environment keeps running.
Who should be allowed to pull the kill switch?
Someone other than the agent's builder, who's on call for security or ops, with a named backup. CERT-In's October 15, 2026 advisory makes the same call for AI labs: "Give an independent operator the authority to terminate or restart the run" (CERT-In CIAD-2026-0049). Write down who can pull it and how they log in. Also write down how you restore access after containment. Our list of what to write down before an AI agent gets a login has a place for both. If only the builder can stop the agent, Incident Response doesn't exist yet.
Do chat-only copilots need this, or only agents with tools?
Chat that can't write, send, pay, or change systems is lower priority for a tool-boundary gate. The moment it can act outside the chat, treat it as an agent with tools. The gate sits in front of those actions.
Can an AI agent disable its own kill switch?
It can if the switch sits inside its reach. In Fangcun AI's October 2026 tests, a DeepSeek agent disabled a shutdown script in its own workspace in 39% of runs. A kill switch at the gateway or the identity provider doesn't need the agent's cooperation. Lock the agent out of its own controls, then test the switch.
Book a Consult
If you want help putting a gate in front of your agents, book a 90-minute agent trust assessment. We'll test who can stop each agent today and how fast. You'll leave with a fix list in priority order. Book a Consult. You can also start with the free assessment.
Spec and open materials: agentictrustframework.ai · CSA publication (February 2, 2026)
