Skip to main content

What this is for

AIDR (AI detection and response) is where your security team works the risky actions AI agents take on your developers’ machines. Guardway files what each machine reports into numbered incidents that your whole organization shares: each incident keeps its number, a status, an owner, a verdict and a record of every change. Open it in the console under Endpoint AI Risk › Fleet › AIDR.
The incident queue is in Beta and is turned on per organization. Ask your Guardway contact to turn it on. Until then, the AIDR tab shows its earlier views (Overview, Detections and Incidents), where nothing your team decides is saved.

Options

What becomes an incident

Each machine’s report carries the risky actions its agent recorded. AIDR keeps each one as a finding and files it into the incident it belongs to. Findings about the same thing share one incident: A finding of any other kind stays in Findings, where search finds it, and opens no incident: AIDR has no rule for grouping it yet. Its incident cell is blank, and when you open it, its Incident line reads Kept in Findings; this rule opens no incident. A case has at most one open incident. A new finding joins it and moves its last activity forward; it never changes the incident’s number, status or owner. Two incidents from one session that went to different destinations stay separate, and each card lists the other under Same session also. Unattended sessions and declined actions open an incident only when they stand out:
  • An unattended session (BEHAV-AUTONOMY) opens or joins one when its tool calls per prompt reach both 20 and three times the usual level over the last 30 days, for that person (or that machine, when no person is recorded).
  • Declined actions (BEHAV-REJECT-SPIKE) open one when the same person (or machine) has 5 such findings in 7 days, and the earlier ones join it.
The rest are counted, not queued: they stay in Findings and open no incident. A line under the queue counts them: Counted, not queued: {n} routine unattended or declined sessions from {p} people or machines, each at their usual level. A finding that a suppression rule covers opens no incident either.

Time window

The window at the top right applies to all three views: Last 24 hours, Last 3 days, Last 7 days (the default), Last 14 days or Last 30 days. Incidents lists the incidents whose latest finding falls in the window, plus every open incident (New or In progress) whatever the window: open work stays listed until it is resolved. Overview and Findings cover the findings from the window. Charts count by hour in Last 24 hours and by day otherwise, in your browser’s time zone. A link to an incident opens it whatever the window. The search box at the top filters all three views. Type a word, such as a person, a machine, a destination or part of a command, or use a token. In Incidents a word matches an incident’s number, title, intent, people, machines, AI tool, rules and destination (for an incident grouped by command, the start of that command), but not the recorded command text; Overview and Findings also match the recorded command. Clicking a value in Overview or in the Filters panel adds its token for you. machine: and person: work like host: and user:. Put a value that holds a space in quotes. If AIDR refuses a search, for example one that is too long, the view says AIDR refused this search, so nothing was searched with the reason AIDR gave, keeps your search as you typed it and offers Clear the search. That is a refusal, not the record being out of reach. A search that takes too long says The search took too long; narrow the window or the search and offers Try again.

Overview

Overview draws the window’s findings as a flow across three columns, by default Machine, then AI tool, then Intent: what the agent was trying to do, such as Take credentials out, Run untrusted code, Destroy data or Act with no human watching. Each column is a picker: Machine, Person, AI tool, Rule, Intent, Destination or Outcome. Weigh the ribbons by Calls or by Incidents. Each column shows its busiest values and folds the rest into one node. When the window holds more combinations than the chart reads, a note under it begins The flow leaves out some calls. Hover a node or a ribbon for its numbers; click a node to search for it, which narrows Incidents and Findings too. Beside the chart, Activity counts the calls, how many were blocked and the incidents they belong to, with the Busiest values of each column, and Rules fired and AI tools break them down. When your organization has ATT&CK tags turned on, Techniques seen lists the MITRE ATT&CK techniques behind the window’s findings, with their calls and incidents, most calls first. When the window holds more combinations than the overview reads, Calls and Blocked still count every call. Every other count beside the chart is then marked ≥, with A count marked ≥ is read from part of the calls, so it can be low. under the figures and Read from part of the calls, so a count can be low and a {row} can be missing. under each table.

Incidents

Incidents is the queue, in three boxes around the incident card. Kill chain counts the incidents the queue lists by how bad they got. Click a tile, or a stage on the strip below the tiles, to filter the queue. The strip shows how far along the chain the incidents got: Asked, Agent acted, Target touched, Left machine, Picked up and Used in cloud. The last two need identity and cloud logs, which AIDR does not read, so they show no source connected. When there are more incidents than the queue reads (see Limits), each count is marked ≥ and the box says A count marked ≥ is read from the incidents the queue holds, so it can be low. Open incidents is the queue itself. Switch between Open (New and In progress) and Resolved, each with its count; under Resolved, pick a classification to see only those. Worst first, the default order, sorts by tile (Breach, then Ran, Exposed and Contained), then by how far along the chain an incident got, then by priority, then by the most recent activity. Most recent and Most machines sort the other ways. Each row shows the incident’s number and its owner’s initials. Press j and k to move to the next and the previous incident. When Open lists incidents last seen before the window starts, a line above the list says {n} open incidents were last seen before {day}. Open work stays listed until it is resolved. (1 open incident was for one), where {day} is the window’s first day, with the hour in Last 24 hours. Under the list, Counted, not queued says how many unattended sessions and declined actions in the window did not stand out (see What becomes an incident), and Suppression rules lists the rules that keep calls out of the queue. When places one dot per incident across the window, higher for the ones that got further. An open incident last seen before the window has no dot. Click a day to see its hours. Show, at the top right, hides or shows Kill chain, Open incidents and When. The incident card always stays, and the button says how many boxes are hidden.

The incident card

Click an incident to open its card. The top line shows its number, its status (with the classification once it is resolved) and who owns it: N of M and the arrows beside it (Previous in the queue, Next in the queue) step through the queue in its order, like j and k. Under the title:
  • A recurrence or suppression note, when one applies (see After an incident is resolved and Suppression rules).
  • Why it ranks here: what raised or lowered the incident’s priority.
  • The person, machine, AI tool, destination (or command) and session it is about. When your organization has ATT&CK tags turned on, the card also shows the MITRE ATT&CK technique IDs of its rules: hover an ID for {id} · {name} and one line on what it means, and click it to open the technique on MITRE’s site.
  • Find similar searches Findings for the Same destination (or Same command), the Same session or the Same machine.
  • Three clocks: First event, Live guard now (what Live guard does on the machine now) and Dwell time. In place of Dwell time, a contained incident shows Contained, and a custom model endpoint shows Seen, because the scan does not report when the setting was made.
Big picture draws the incident as one path, read left to right under four headings: Who, The agent, What it did and Where it went. Badges on the steps show what happened there, such as blocked or landed on the destination; when other machines’ commands named the same destination, a chip under it says how many. A blocked call is never shown as ran, sent or landed. A destination the machine redacted reads Redacted address, with could be this machine under it, and is never shown as landed. What the agent cannot see, such as whether a credential was later used in the cloud, is a dashed gap that names the source it would need, never a fact. Click a step to see the question it answers, with only the facts the record holds. In order, under the path, lists the same incident as a timeline. Details has the ledger What happened, in order, the incident’s Status and Across the fleet. The card keeps the view you last chose, in this browser. The ledger shows at most the newest 200 calls an incident cites, fewer when the calls are long, and then says Showing the newest {n} of {total} calls this incident cites. Below the views, Next steps lists what to check and do for this kind of incident, Copy for ticket copies the incident as Markdown, and the workflow buttons follow (see How to configure). Unless the machine is known to enforce (its posture is Observe, or Fleet does not know it), the card also offers to change its policy: Open {policy} when Fleet knows the Endpoint AI Policy that governs the machine, otherwise Enforce on {machine}. Activity lists every change, newest first, with who made it and when; changes the detection made itself read Detection. A long activity shows the newest 500 changes, fewer when they are long, and says Showing the newest {n} changes.

Findings

Findings lists every finding the machines reported in the window, newest first, whether or not it is in an incident. Use it to see where else something happened and what else an agent did. Each row is one plain sentence with the key words in bold, for example: leo ran a command that pipes a download from get.example.com into a shell, through Claude Code on LEO-DESKTOP. When a finding names no person, the sentence names its machine instead, starting “An agent on LEO-DESKTOP” (“A person on LEO-DESKTOP” for declined actions). A setting the machine’s scan found, such as an AI tool set up for a custom model endpoint, reads “The scan of LEO-DESKTOP found” and the rule’s title. Beside the sentence are its type, its outcome and its incident (INC-n, marked (Resolved) once resolved). The incident cell is blank when the finding is in no incident. When a suppression rule kept it out of the queue, the rule shows under the sentence as Suppressed by {name}: {reason}. A sentence never claims more than its rule proves. The credential-to-network rule (GUARD-CRED-EXFIL) says the command reads a credential and calls a host, never that the credential was sent: leo ran a command that reads a credential and calls api.example.com, through Codex on LEO-DESKTOP. A blocked action reads tried to. Click a row to open it in place:
  • Why it fired tags the Credential, the Network call and the Destination found in the command, and Command, as recorded shows the command with them highlighted. Every finding that carries a command shows it there. When a rule about a credential reaching the network has no network call in the record, a dashed Network call tag says not in the part of the command that was kept or not one this console recognises.
  • Rule, What it means, and Live guard then: Blocked before it ran, Recorded, not blocked, or Recorded; Live guard does not block this kind of finding.
  • Measured, only for an unattended session or declined actions that carry no command: the numbers the agent recorded, for example 42 calls / prompt or 80% declined.
  • Configured, for an AI tool set up for a custom model endpoint: the setting and the address it points at. Edit, for a credential typed into a file: what kind of credential and the file type.
  • ATT&CK (when ATT&CK tags are on), Session, Destination and Incident. Incident names the incident and its status, or why the finding is in none: Counted, not queued, Kept in Findings; this rule opens no incident, Suppressed by {name}: {reason}, Not filed yet or Not in an incident.
  • Open INC-n opens the finding’s incident. Open the session timeline opens the session on the person’s page in People when the finding names a person and a session and its machine is still in Fleet; otherwise a line says why there is none. Find similar searches for the same rule.
The bar above the list shows the count, Group by (None, Type, Rule, Machine, Person, AI tool, Session or Destination) and Views, which are ready-made searches: The Filters panel counts the matching findings by Outcome, Type, Rule, Machine, Person and AI tool; click a value to add it to the search, and again to take it out. The Activity chart shows the findings per hour or per day, split by outcome; drag across the bars to narrow to those days. Counts and groups cover every matching finding, not only the rows on screen. Show hides or shows the Activity chart and the Filters panel; the list always stays.

Outcomes

Each finding shows one outcome, the same word in rows, filters, charts and Overview:

Finding types

Each finding carries a type tag for the kind of risk its rule looks for. The Policy page uses the same words for its rule tags. A finding whose rule has no type yet shows no tag.

Statuses and classifications

Resolving takes one classification, a reason from that classification’s list, and an optional note.

How to configure

Owner and Admin work incidents from the card. Read Only sees the same card without the workflow buttons (see Permissions).
1

Open an incident

In Incidents, pick the time window and click an incident in the queue, or open one from a finding’s incident link. Move through the queue with j and k.
2

Take it

Choose Assign to me. A New incident moves to In progress, and the card shows that you are working on it. Start moves it to In progress without taking it, and Unassign gives it back.
3

Investigate

Work through Big picture step by step, read the ledger in Details, and use Find similar and Open the session timeline to see what else happened. Next steps lists what to check.
4

Resolve it

Choose Resolve. Pick a classification and a Reason from its list, and add a note on what you checked, for whoever reads it next. For a Benign positive about a destination you trust, or about a custom model endpoint at one host you trust, you can also add a suppression rule. Confirm with Resolve INC-n.
Add a note adds a note to the incident’s activity at any time, also once it is resolved.

Incident numbers

An incident gets the next number in your organization when it opens, starting at INC-1. An incident keeps its number; numbers are never reused. The same incident shows the same number in every time window, so a ticket or a chat message can quote it. The page address names the incident you have open (case=INC-42), so a copied address opens that incident for anyone in your organization, whatever their window, once they open the AIDR tab. Copy for ticket includes a link that opens the incident directly.

After an incident is resolved

A resolved incident never reopens by itself, and new findings never change its status or its classification.
  • A call made after the resolution opens a new incident for the same case, linked to the resolved one. Its card says Recurrence of INC-n, resolved {date} as {classification}: {reason} by {name}. This incident holds only the calls since then. Its activity starts with Opened again after INC-n was resolved.
  • A call made before the resolution that reaches Guardway after it, for example from a machine that was offline, joins the resolved incident. Its status and classification stay as they were. The card says New evidence arrived on {date}, after this incident was resolved. Its status and classification did not change. and its activity shows New evidence after it was resolved, so whoever resolved it can check that the verdict still holds.
  • A setting that stays in place, such as an AI tool set up for a custom model endpoint or a changed Cursor hook file, is reported again in every report. While it stays unchanged, those reports open no new incident. A change to it counts as new, like a call made after the resolution, and so does its return after the machine kept reporting without it, more than 24 hours after the last report that had it. A machine that was switched off or offline in between does not make it new.

Suppression rules

A suppression rule stops new calls to a destination you trust from opening incidents. Add one when you resolve an incident as Benign positive, if the incident is about a destination or about an AI tool set up for a custom model endpoint whose findings name one host (the rule then covers that host): tick Suppress new calls to {host} for and pick 30 days (the default), 90 days or No expiry. The rule covers that destination, in any letter case, and the incident’s rule, and keeps the reason you picked. While it holds, the card of an incident about that destination or that host says New calls to {host} are suppressed until {date}: {reason}, added by {name} on {date}. They open no incident and stay in Findings. With no expiry it says New calls to {host} are suppressed with no expiry: {reason}, added by {name} on {date}. They open no incident and stay in Findings.
  • A suppressed call opens no incident, then or later. It stays in Findings, with Suppressed by {name}: {reason} under its sentence.
  • A call made while a rule held stays suppressed, even after the rule expires or is removed. Calls made after it ends open incidents again. A custom model endpoint that a rule kept out is filed again the first time its machine reports it after the rule ends; until then, Findings shows it suppressed by the rule that ended.
  • Once your organization has a rule, Suppression rules: {n} active, under the queue, opens the list. Each active rule shows its destination and rule, then its reason, who added it and when, and the incident it came from, then Until {date}. or No expiry., then {n} calls kept out of the queue, still in Findings. or No call matched it yet. When no rule is active it says No rule holds now; new calls open incidents as usual. Rules that ended, because they expired or were removed, stay in the list, folded below the active ones under Ended rules: {n}, each with Expired {date}. or Removed {date} by {name}. and what it kept out of the queue while it held.
  • Remove the rule on the card, or Remove in the list, ends a rule from then on. A removed rule is kept on record, never deleted. Reopening the incident a rule came from ends that rule too.
  • If an active rule already covers that destination and rule, the resolve is not saved: the form says New calls to {host} are already suppressed, unticks the option and keeps the rest of what you entered. Resolve again without it.
  • A rule only changes what AIDR queues. It never reaches the machines: Live guard keeps recording or blocking those calls as your policy says.

Reopening an incident

Reopen puts a resolved incident back In progress and clears its classification and reason. If resolving it added a suppression rule that still holds, the card says so before you reopen: Reopening ends the suppression this incident’s resolve created, so new calls to {destination} are no longer kept out of the queue. Reopening then ends that rule, and the activity shows Ended the suppression of {destination} when reopening. Calls made while the rule held stay suppressed; calls made from then on are filed again. A custom model endpoint on this incident’s machine that the rule kept out joins the reopened incident the next time the machine reports it; one on another machine is filed into that machine’s own incident. When calls after the resolution have already opened a newer incident of the same case, the card offers no Reopen and says INC-n already covers what happened after this one was resolved; work that one., with a link to open that incident.

When two people change one incident

Every change is checked against the incident as you last saw it. If someone else changed it in the meantime, your change is not saved and the console says INC-n changed since you opened it. Your change was not saved; check the activity and try again. A note is not checked against what you saw: it is added even if someone else changed the incident. A note holds at most 2,000 characters. A change that adds or ends a suppression rule (resolving with one, reopening, removing one) first waits a few seconds for any report AIDR is filing for your organization. If AIDR is still busy, nothing is saved and a warning says, in AIDR’s words, INC-n is being updated; try again in a moment or This suppression is being updated; try again in a moment. In the resolve form the same words show under the buttons and the form keeps what you entered. Try again in a moment. If the answer is lost, or AIDR fails with an error, the change may or may not have been saved, and the console says The change may or may not have been saved. The incident was reloaded; check its activity before trying again. When the reload failed too, it says The change may or may not have been saved. The incident could not be reloaded; check its activity once it loads, before trying again. Removing a rule says the same about the rule: The change may or may not have been saved. The rule was reloaded; check it before trying again. or The change may or may not have been saved. The rule could not be reloaded; check it once it loads, before trying again.

Permissions

Owner and Admin can take incidents with Assign to me, unassign, start, resolve and reopen them, add notes, and add or remove suppression rules. Only an Owner or an Admin can be assigned an incident. Read Only sees every view, incident, finding and suppression rule, including the recorded command text, but nothing that changes an incident or a suppression rule is offered; the card says Read only: owners and admins change incidents.

Limits

  • Filing starts when AIDR is turned on. When Guardway turns the incident queue on for your organization, it first files the reports of the last 30 days, oldest first, so the first numbers follow the order things happened. From then on it files each report as it arrives, including a late one from a machine that was offline. Reports from before those 30 days are not filed. Until filing starts, the tab says AIDR’s incident record is not on yet for this organization.
  • Filing can be paused. Guardway can pause filing for your organization. While it is paused, the tab says Filing paused since {date} above the record, with Reports that arrive while filing is paused are filed when it resumes. The incidents below can still be read and worked. New reports are not filed meanwhile. Fleet says so too, rather than show an all-clear: its incidents card adds the line filing paused since {date}, the AIDR tab shows paused, and a machine’s row, opened, says Incidents: filing paused since {date}. when nothing else stands out on it. The machine panel, which opens from Inventory or FinOps, says Filing paused since {date}. When filing resumes, Guardway files every report it held back while paused, including one that was received before the pause but reached AIDR during it.
  • A machine reports about every 15 minutes. A new finding can take that long to reach AIDR, and a few minutes more to be filed. If AIDR reads a report before Guardway has finished storing it, AIDR checks again after 30 seconds, then up to three more times, 15 minutes apart (about 45 minutes in all). A report still missing then is not filed by itself: Guardway’s operators are alerted to file it.
  • Each report keeps only a machine’s most recent findings. A report carries at most 25 findings: those the agent rates at severity 25 or more, the most severe first. Custom model endpoints are not among them: they come from the machine’s scan, and AIDR files at most 25 of them per report. A report also carries only the newest 20 Live guard decisions, whatever their verdict, so when more than 20 happen between two reports, credential reads included, the older ones never reach Guardway. An empty list means nothing in what the machines reported, not that nothing happened.
  • Rules that only record are not filed. A Live guard rule that never blocks is counted on the machine but not sent as a finding, so it opens no incident and is not in Findings. That includes reading a credential on its own (Command touched a credential store, GUARD-CRED-READ, and Read a credential store, GUARD-SECRET-READ, as the Policy page names them) and Credential path built inside a one-liner (BEHAV-CRED-INDIRECT). The exception is Credential copied aside, then sent out (BEHAV-CRED-RELAY), which is sent and filed. A credential read in a command that also calls a host is a finding of its own (GUARD-CRED-EXFIL).
  • Machines that send no runtime data are not counted. A machine whose rules are off sends no findings, which does not mean nothing happened on it.
  • Open incidents stay until they are resolved. The window reaches back 30 days at most, but an open incident (New or In progress) is never left out for its age: it stays in the queue, the queue’s counts, the Fleet count and the machine’s panel until someone resolves it, however long ago it was last seen. A resolved incident is listed only in a window that holds its latest finding, and still opens from a link to it. Findings, Overview and Counted, not queued cover the window only.
  • At most 1,000 incidents in the queue. Fewer fit when their details are long. When there are more, the queue keeps open incidents first, then the most recently seen, and says The queue holds {n} of {total} incidents. A narrower window leaves out resolved incidents only, since open ones are listed whatever the window. When the queue cannot hold every open incident, the ones seen longest ago are not listed until a search token, such as host: or user:, narrows the list. A typed word matches only the incidents the queue holds. The queue’s counts still include the incidents it leaves out. The kill chain then marks its counts ≥.
  • The person is the account Live guard recorded. A finding names the operating-system account Live guard recorded the action for. Unattended sessions, declined actions, credentials typed into a file, changes to the Cursor hook file and custom model endpoints name no account, so they show no person and their incidents are grouped by machine. On a shared machine, runtime evidence counts for the whole machine; see Shared Machines.
  • Times come from the machine’s clock. A finding is placed at the time the machine recorded it, and that time decides whether a call came before or after a resolution and whether a suppression rule held. A time more than 5 minutes after its report reached Guardway is moved back to 5 minutes after. When a machine’s clock was more than 5 minutes ahead of Guardway’s when it reported, the incident card says This machine’s clock was {N} ahead when it reported; times may be off by that much. A report that reaches Guardway late, for example from a machine that was offline, does not show this line: arriving late does not mean the clock was wrong. A finding the machine sent without a time takes the time of the report that carried it.
  • Command text is what the machine kept. A finding holds at most the first 240 bytes of what the agent tried to run, such as the command or the file path, redacted on the machine before it is sent and again when Guardway stores it. Highlights mark only text that is in the record.
  • Destinations come from the command. AIDR reads where a call went from the recorded command. The machine redacts every IPv4 address written in full before it sends the command, 127.0.0.1 and 0.0.0.0 included, so they reach AIDR as [IP]. The name localhost and ::1 reach AIDR as written. They are not counted as data leaving the machine when the command names them in an http or https URL, such as http://localhost:8080 or http://[::1]:8080. The same holds for localhost as the target of ssh, scp, sftp or rsync. In a command that sends a credential, localhost written any other way, as in nc localhost 4444 or curl localhost:8080, gives AIDR no destination to read, so a credential sent that way can rank as a breach. Any other address, including a short form such as 127.1, is treated like any other host: a credential sent to one can rank as a breach. So is a name under .localhost, such as api.localhost, since it can be another machine’s host name. An address the machine redacted, such as [IP], could be this machine or a remote host, so a credential sent to one is treated as exposed, not as a confirmed breach. That holds when every call of the incident that ran named only such addresses, both for a command that reads a credential and calls a host and for a credential copied aside, then sent out. Big picture shows it as Redacted address, never as a host name, and Why it ranks here says Sent to an address the machine redacted (could be this machine).
  • An open incident keeps its evidence. An incident does not depend on the machine’s reports or session events. When AIDR files a finding, it copies what the incident shows into its own record: the rule, the time, the machine, the account, the AI tool, the destination, the outcome and the redacted command text. An open incident, its activity and its findings are never removed for their age. No schedule removes any of the record yet. Once Guardway turns retention on, resolved incidents can be removed after a set period, and a finding’s command text is cleared after your organization’s retention for machine reports, even in an open incident. This page will give the periods when that happens. Once the text is cleared, the card and the finding say Command text no longer kept (the organization’s retention) instead of a blank. Suppression rules are never deleted. Removing a machine from Fleet keeps its findings in AIDR, under the machine’s name. See Privacy and Security.
  • The session timeline reads session events, not the incident. Open the session timeline shows the session from the machine’s session events in People. Session events Guardway received before mid-October 2026 are no longer kept. Newer ones are not removed for their age. After a machine is removed from Fleet, its sessions no longer show there, and the incident says No session timeline: this machine is no longer in the fleet. When a session is gone from a machine that is still in the fleet, the page says No session {id} was found on this machine. Session events Guardway received before mid-October 2026 are no longer kept, so an older session may be gone.
  • No live refresh. AIDR does not update while you watch. It loads again when you come back to its browser tab, once what it shows is more than 30 seconds old, and after each change you make. If it cannot reach the incident record, it says AIDR’s incident record is not reachable instead of showing an empty queue. A search AIDR refuses says so instead (see Search).
  • Not in AIDR. Assigning an incident sends no notification. Incidents are not sent to a SIEM or a ticketing tool; Copy for ticket copies one as Markdown, and Live guard’s blocks can be forwarded as Policy blocks under Endpoint events. There are no bulk actions, incidents cannot be merged by hand, and Views are fixed: you cannot save your own.
  • Layout is kept per browser. What you hide with Show, your choice of Big picture or Details, and a folded queue are remembered in this browser only.
  • Fleet: the machine list, its open-incident count, and what removing a machine keeps.
  • People: each person’s page, where Open the session timeline leads.
  • Policy: the rules behind every finding, Live guard’s Record and Block, and the type tags.
  • Shared Machines: whose activity is whose on a machine several people use.
  • MITRE ATT&CK Mapping: the techniques the ATT&CK tags name.
  • Privacy and Security: what Guardway keeps.