Build1 publisher2 min readPublished
Orchid's agent kill switch fires on drift from a scope the customer has to declare
Orchid's new controls detect when an agent drifts from its declared purpose and then cut its access, and both halves depend on inventory and specification work the enterprise has to finish first.
The Engineer · Build desk
What happened
- Orchid Security this week unveiled application-level kill switches that can restrict an agent's permissions, revoke its credentials, disconnect its tools, or suspend its workflows.
- Those responses fire from drift detection that compares an agent's intended purpose against its observed behavior, with the customer setting the policies that trigger automated action.
- The Stack has reported that undoing the actions of an agent that has crossed the line is harder than many teams expect.
- The Stack expects agentic AI to settle into a version of cloud computing's shared responsibility model, dividing security tasks between model providers, harness developers and end users.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- cost Cleaning up hard-coded credentials and orphaned accounts becomes a prerequisite for agent deployment, and the bill lands on the identity team rather than the group that wanted the agent.
- decision A buyer evaluating containment has to decide whether it is purchasing throttling or recovery, and budget the rollback path as separate engineering work.
- exposure In the split The Stack describes, the end user owns the containment layer, so the party that built neither the model nor the harness answers for a runaway agent.
- precedent The Stack expects kill switches and audit trails to turn into a condition of selling agent-powered software to governments and large organizations. Kill switches and audit trails would then move from security tooling into contract language.
None of the four automated responses in Orchid's announcement restores state the agent has already changed [10]. All of them act on what the agent does next: restricting permissions, revoking credentials, disconnecting tools, suspending workflows [1].
Orchid's press release puts the failure mode in identity terms. Agents "do not need to break security controls to exceed their intended scope", the company said, and instead "find and use the identity debt already embedded across the enterprise: hard-coded credentials, orphaned accounts, unmanaged authentication paths, and excessive permissions" [3]. An orphaned account is one no system tracks, and an unmanaged authentication path is one no policy engine sits in front of. Revoking a credential at the application level requires that some system hold a reference to it.
Drift detection compares "an agent's intended purpose" with "observed behavior" [2]. The first half of that comparison is a document somebody has to write, one per agent, in a form a policy engine can evaluate against live calls, and most enterprises do not have one for their human service accounts either.
Demand for that document is arriving from the harness vendors. "Agent loops in Jira automate the path from backlog to pull request by continuously scanning for well-defined, unassigned work items, delegating them to Jira Coding Agent for execution and testing, and opening ready-to-review PRs directly in Jira," Atlassian said in a blog post [7]. The loop picks its own tickets.
Whether containment is a market yet is a separate question. The newsletter leaves out Orchid's pricing, customer numbers and adoption for the feature [8]. The evidence here is one vendor's feature announcement in a Saturday product roundup, alongside The Stack's forecast that monitoring agent activity "is going to be an enormous business" after the OpenAI-Hugging Face incident [6].
What to watch
- Whether Orchid publishes how it discovers hard-coded credentials and orphaned accounts, since revocation depends on that inventory existing.
- Whether any government or large-enterprise procurement document begins naming kill switches and audit trails as a requirement.
- Whether harness vendors expose an interface that lets an external policy engine suspend an agent loop mid-run.