PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteiTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
If an AI agent is doing something you did not authorize, stop its run immediately, prevent it from issuing more actions, and inspect what it has already done. Stopping is not rollback: a tool action may have completed before the stop took effect. To reduce the chance of a repeat, limit the agent’s permissions and access, and put an independent authorization check in front of consequential actions.
What to do while the agent is running
Use this sequence whether you are using an agent in an app or operating one built into a workflow. Product controls differ; for a developer-operated agent, stopping the run should also stop new tool dispatches.
- Stop the run. Use the product’s stop control. If you operate the agent, halt dispatch at the orchestration or tool-execution boundary so it cannot issue another operation. OpenAI advises users to stop tasks immediately if something seems suspicious in its prompt-injection guidance.
- Do not automatically retry. A retry can repeat the same action or create another side effect. OpenAI’s API guidance on misalignment monitoring specifically says not to automatically retry a workflow blocked by the monitor.
- Check what actually happened. Review the agent’s tool calls and outputs, the resources they touched, and the records in the connected application. An alert is a reason to investigate, not proof that a particular action was unauthorized—or that no action completed.
- Preserve relevant records. Keep the request and response IDs, tool calls, outputs, and application records needed to reconstruct what the agent could see and do. Follow your organization’s rules for handling sensitive information in logs.
- Contain the implicated access. If you administer the system, temporarily disable or narrow the implicated tool, connector, or credential while you investigate. The exact procedure depends on the platform and connected service; there is no single universal revocation path.
- Assess and repair effects deliberately. Determine what changed and whether the affected application can safely reverse it. A stop signal does not undo an email, payment, file change, or other completed action; there is no universal rollback for different tool effects.
If you are an ordinary user without permission to change connectors or execution policies, stop the run, review and report what happened, and ask the service or workspace owner to restrict access or add an approval gate.
Free tools Windows power users keep installed
One-click scans. No signup required.
Why an agent can act outside your intent
An unwanted action does not necessarily mean the user gave a malicious instruction. An agent may combine trusted directions with untrusted material from a webpage, email, document, or tool response. NIST describes this kind of agent hijacking: malicious instructions embedded in data the agent consumes can steer it toward an unintended action. Broad requests such as “handle everything” also leave room for interpretation.
#1 Best Overall
Risk increases when the agent can turn that interpretation into a consequential action without another system checking whether the specific action is authorized. OpenAI’s March 11, 2026 article recommends thinking about both the source that can influence the agent and the sink that can cause harm—for example, untrusted text paired with the ability to transmit information, follow a link, or invoke a tool. Reducing either side can limit risk; relying only on detecting malicious input leaves the action capability intact.
Which controls help, and when
Controls work best in layers. A prompt can clarify intent, but an independent execution check can deny an action even if the model is confused. Monitoring and stop controls help during a run, but they may not act before an operation completes.
| Control | When and where it acts | What it can limit | Important trade-off |
|---|---|---|---|
| Narrow instructions and task scope | Before execution; in the request or agent configuration | Ambiguity about permitted targets and actions | Clarifies intent but does not technically remove a tool’s capability. |
| Least privilege and restricted data access | Before execution; in permissions, connectors, and identity configuration | Resources and operations available to the agent | May limit what the agent can accomplish; permissions must match the real task. |
| Independent authorization and human approval | At execution; in the tool runtime or an external policy layer | Specific actions that lack valid authorization | Review adds delay. Approval must describe the exact action and parameters. |
| Monitoring and stop controls | While or after actions occur; in the product or orchestration layer | Further activity after detection | Detection can be late, can miss issues, or can flag legitimate activity. |
| Logging and task-specific testing | Across operation and before re-enabling autonomy | Gaps in investigation and untested failure paths | Tests need to reflect actual tools and tasks; logs need suitable data protections. |
Make a repeat less likely
Define a bounded task
State the goal, allowed targets, and permitted actions. Replace open-ended directions with a specific request where possible. A clear instruction reduces interpretive latitude, but it is not a substitute for technical controls.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Reduce available data and capabilities
Enable only the applications and data the task needs. Use logged-out or read-only access when it is sufficient, and separate data or memory across users and sessions when appropriate. Give the agent only the tools, operations, and resource scope it requires. As OWASP puts it in its AI Agent Security Cheat Sheet, “Grant agents the minimum tools required for their specific task.”
Rank #3
A natural-language rule such as “do not delete files” does not remove delete capability if the connected tool still permits deletion. Permissions should enforce the boundary rather than relying on the model to obey the sentence.
Authorize the exact action outside the model
Place an authorization check in the code, policy service, or identity layer that executes tool calls. It should evaluate the acting identity, tool, target, parameters, and required approval. Fail closed when the tool is unknown or required authorization is missing. If a parameter changes, obtain a new approval for the changed action; a general approval flag is not authorization for different parameters or a later operation.
Require review for consequential actions
Require a person to review the exact proposed action before the agent can send, purchase, delete, modify a system, or expose sensitive information. The reviewer should be able to see what will happen, to which target, and with which parameters. This is more meaningful than approving a broad task in advance.
Set operational limits and keep useful records
Set hard limits on retries, chain depth, tokens, cost, and other relevant execution budgets so repeated calls cannot compound effects indefinitely. Keep enough records to investigate behavior and detect anomalies, with protections appropriate to sensitive content in logs.
Best Value
Test the actual workflow before restoring autonomy
Test the same tools, data sources, and actions the agent will use—not only a generic prompt-injection benchmark. Include indirect prompt-injection attempts through the agent’s own websites, documents, APIs, and tool responses. Test the prior failure path, repeat attempts, and confirm the block occurs at the execution boundary. Also verify that the intended task still works.
Retest when the model, tools, permissions, or surrounding system changes. Aggregate scores can conceal a weakness in one important task, so examine results by task and action as well as overall. NIST’s January 2025 CAISI evaluation article reports that the tested agents were frequently induced to follow malicious instructions across three added risk areas. Those findings concern particular historical model versions and test setups, not a measurement of today’s agents or a universal attack rate.
What monitoring can—and cannot—guarantee
Monitoring can help identify suspicious behavior, but it is not a guarantee that every problematic action will be caught in time. OpenAI says its user guidance may not prevent every prompt injection. Its API documentation explains that monitoring is asynchronous, may miss issues or flag legitimate activity, and may identify a concern only after an action has completed. Treat a monitor as one layer, not as authorization or rollback.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
OpenAI’s March 11, 2026 article reports that one prompt-injection example from 2025, described by external security researchers under a specified user prompt, worked 50% of the time. That is a result for that particular reported test—not a general prompt-injection success rate or a forecast for another agent, model, task, or date. No broader comparable success-rate figure across AI agents is established by the cited sources.
What users and system owners can change
A user can stop a run and report an unwanted action, but may not control the agent’s connectors, permissions, or execution policy. Those changes generally require the service or workspace owner. Controls also vary by product: Anthropic, for example, describes Claude Code as read-only by default with human approval before modifications in its framework for developing safe and trustworthy agents; that is a vendor-specific example, not a default shared by all agents. For ChatGPT agent, consult the current Help Center guidance for product controls.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
