We Killed the Agent, but It Keeps Moving

Cyber Explorers, stopping an agent is not the same as ending its influence.
A model can stop producing new output while its memory, permissions, scheduled work, copied recommendations, and machine-shaped operating rules remain distributed across other systems. That is the distinction this week’s security guidance makes harder to ignore.
The development
NIST’s November 9 initial public draft, SP 2400-102e2026, adds a dedicated taxonomy for attacks on agentic systems. Its long-horizon section describes persistent software actors that accumulate state, memory, tool access, permissions, and operational trust across sessions. NIST also says ordinary reboot-and-restore assumptions can fail when valuable learned context, poisoned adaptation, or durable state cannot be cleanly rolled back. The draft remains open for public comment through December 18, 2026.
OWASP’s Securing Agentic Applications V2, released November 10, translates agentic threats into runtime controls: interception hooks, memory-write checks, provenance snapshots, time-to-live rules, and graduated circuit breakers from warning through kill. OWASP is also candid that framework coverage is uneven. Memory, knowledge-retrieval, sub-agent, compaction, and skill-lifecycle hooks remain sparse or inconsistent.
OpenAI’s Control Room, published November 9, supplies a concrete implementation example. It can halt new model output across research inference workloads within 35 seconds of a stop decision and fails closed if the control service is unavailable. OpenAI explicitly notes that the workload process may remain alive even when further model output is denied.
Shutdown is a control event—not a governance conclusion
These sources converge on an important operational point: a kill switch answers whether new execution may continue. It does not, by itself, answer what must be revoked, invalidated, quarantined, or re-examined after the stop.
Cyber Explorer working construct: a defensible agent shutdown should close five different surfaces. This is an analytical extension of the cited material, not a NIST or OWASP requirement.
Execution closure: Can the agent, its sub-agents, background jobs, or alternate inference paths still act?
Authority closure: Have delegated credentials, tool permissions, approval tokens, and human-on-behalf-of grants been revoked or expired?
State closure: Which memories, caches, context summaries, embeddings, and queued tasks remain active, and which must be quarantined or rebuilt?
Derivative closure: Which downstream decisions, tickets, policies, configurations, or reports still rely on the agent’s earlier judgment?
Evidence closure: Was enough provenance preserved to reconstruct why the agent acted without retaining sensitive reasoning-layer data indefinitely?
Why the distinction matters
Consider a hypothetical contractor using an internal agent to review configuration evidence. The agent is stopped after anomalous behavior. Its inference endpoint is no longer reachable—but its risk scores have already populated a dashboard, a scheduled remediation job is still queued, an analyst approved a ticket based on its summary, and a compacted memory record now informs the replacement agent.
No new output is being generated. Yet machine-mediated authority is still moving through the environment.
Traditional controls can address pieces of this problem: account revocation, job cancellation, log preservation, configuration rollback, records management, and incident response. What remains weakly explicit across NIST AI RMF, NIST SP 800-53, and much of current agent guidance is the requirement to trace and close the full chain between machine inference, retained state, delegated authority, and downstream human adoption.
A practical closure test
Inventory every active path. Map the agent, sub-agents, tools, service identities, background processes, callbacks, queues, memory stores, and downstream consumers.
Bind authority to current purpose. Revalidate who authorized each surviving permission and whether that authorization still applies after the stop event.
Quarantine before deletion. Preserve the minimum evidence needed for reconstruction, then isolate suspect state so a replacement agent cannot silently inherit it.
Mark derived outputs. Tag recommendations and decisions produced by the stopped agent as pending revalidation, especially where a human merely approved machine-framed evidence.
Declare closure criteria in advance. A shutdown playbook should define when execution, authority, state, derivative products, and evidence are each considered closed—not just when the model stops responding.
Evidence boundaries
NIST SP 2400-102e2026 is an initial public draft, not a final standard or compliance mandate. OWASP’s guide is technical builder-and-defender guidance and explicitly leaves broader governance to a separate effort. OpenAI describes its own research infrastructure, not a generally available control plane or proof that every external agent deployment can be stopped the same way.
The cited sources establish that long-running agents accumulate durable state and trust, that runtime interruption needs hard enforcement points, and that stopping inference can leave a workload alive. The five-surface closure model above is Cyber Explorer analysis intended to make those implications operational.
What Cyber Explorers should do now
Add residual-state and derivative-output checks to agent incident-response exercises.
Require agent identities and delegated permissions to have explicit owners, expiration, and revocation paths.
Test whether shutdown reaches sub-agents, external tools, queues, scheduled jobs, memory stores, and alternate model endpoints.
Preserve provenance before rollback, then revalidate downstream decisions that depended on the stopped agent.
Use the NIST comment window to ask for clearer terminology around authority revocation, residual machine judgment, and lifecycle closure.
Sources
NIST SP 2400-102e2026, initial public draft — published November 9, 2026
OWASP Securing Agentic Applications V2 — released November 10, 2026
OpenAI Control Room — published November 9, 2026
Disclosure: AI assisted with research, drafting, and production under Cyber Explorer's standing editorial instructions. Allen Westley retains editorial responsibility. Cyber Explorer's independent research does not represent any employer, customer, or government agency.





Comments