AI Pentest Assisted

Introduction

AI Pentest Assisted is a mode of AI Pentest where a human analyst supervises the run live. The AI agent still runs on its own — it never waits for approval — but the analyst watches what it does in a dedicated war room, steers it with directives, parks or aborts steps, asks it questions, and confirms or dismisses every finding before it counts as a vulnerability.

Use Assisted when you want the speed and coverage of the autonomous agent together with human judgment on what gets reported:

• Proof before promotion. Each finding arrives as a draft. It only becomes a vulnerability after an analyst reviews the evidence and confirms it. False positives are dismissed before they reach your vulnerability list.
• The AI acts, the analyst governs. Hold, release, stop a step, or redirect the agent at any time. Controls always stay reachable and take effect visibly.
• Show the work. The live activity feed, phase progress, and attack graph show what the agent tried, where it is, and why it is stuck — not only what it found.

The analyst can be someone from your own security team or a Conviso pentester, depending on your contract.

How it works

Assisted is a third artifact type, next to Single and Recurring. Everything you already know about artifacts — Application, Scope, Safety, Authentication, Code & Documentation, Assessment Settings — works the same way. What changes is who is involved in each run and how findings are promoted.

	Single	Recurring	Assisted
Who starts a run	A user (on demand) or a one-time schedule	The scheduler, on a cadence	A user, on demand
Scheduling	Optional, once	Weekly / monthly / quarterly	Not available — an assisted run is started by someone who intends to watch it
Human during the run	No	No	Yes — lead analyst and participants in the war room
How findings are created	Directly as vulnerabilities	Directly as vulnerabilities	As drafts, confirmed or dismissed by an analyst
Credits	Cost model	Cost model	Same cost model — an assisted run costs the same as an autonomous one

Prerequisites

Assisted runs share every prerequisite of AI Pentest (permissions, credits, an Application with FQDN assets, and optional Secrets, repositories, and documentation). In addition:

Requirement	Needed for
At least one analyst with access to the company	The Lead analyst and participants must be platform users with access to the company that owns the artifact.
AI Pentest update permission	Acting in the war room (controls, directives, finding decisions). Users with read access only can watch.
A browser session kept open during the run	The war room is a live screen. The run does not depend on it, but supervision does.
Traffic allowed from the pentest IP	Shown on the Scope card, same as any AI Pentest run.

Creating an assisted artifact

1. Open AI Pentest in the side menu.
2. Click New (top right of the artifact table) and choose Create assisted.
3. Fill in the cards as you would for any artifact — see Configuring an artifact. The Type shows Assisted and is locked.
4. Fill in the Assisted setup card (it replaces the Scheduling card).
5. Save.

Assisted setup card

The card opens with a reminder of how the mode behaves: "The AI still runs on its own" — it does not wait for approval, and you can hold the next step, redirect scope, validate exploitation, or dismiss a finding at any time.

Field	Description
Lead analyst (required)	The analyst who owns the run. They are the primary person supervising the war room.
Others who can join	Additional users allowed into the war room. Everyone listed here (plus the lead analyst) is a participant.
Ping the analyst when	Which moments send a notification. See Notifications.

The ping options are:

• the run starts
• a Critical is found
• the AI is about to write to the target
• the AI gets stuck
• every phase change

note

The platform checks that the lead analyst and participants have access to the company before saving. If you later switch an artifact away from Assisted, its analyst assignments are cleared.

Starting a run and entering the war room

Open the artifact and click Run AI Pentest. The confirmation dialog shows how many credits the run consumes. Confirm to queue the execution, exactly like any other AI Pentest run (see Running an AI Pentest).

Each execution card of an assisted artifact has an Enter war room button. Click it to open the war room for that run. The war room becomes available once the run starts (before that you see "This execution has no war room available yet.") and stays available after the run ends, with whatever still needs a decision.

tip

Only participants (the lead analyst and the users in Others who can join) can open the war room. Anyone else gets "Could not open the war room. Check your connection and that you are a participant, then try again."

The war room

The war room is a full-screen view: the side menu and global header are hidden so the run gets the whole viewport. Use Exit war room (top left) to go back to the artifact. The breadcrumb shows the artifact, the run, and the war room.

The screen has three areas, each scrolling on its own. On smaller screens they become tabs.

Area	What it shows
Header and control strip	Run number, elapsed time, who is Watching, the run state (Live, Pause requested, Held, Ended, Failed, Cancelled), the current step, and the run controls. Alerts appear here.
Run phase	A progress bar with five phases: Recon, Discovery, Exploitation, Post, Report.
What the AI is doing	The live activity feed. Filter it by All events, Only findings, or Only mine. The feed opens on the most recent events; click Load older events to page back to the start of the run.
Session	One shared thread with assistant questions and answers, directives, and control actions in chronological order, plus the session composer.
Needs your review	The review queue of draft findings and the detail panel. See Reviewing findings.

Everything in the room is shared: every participant sees the same feed, thread, queue, and decisions in real time.

Controlling the run

The control strip has three controls. They change what the agent does next; they never stop the whole run (to stop the run entirely, use Cancel run on the artifact's execution card).

Control	What happens
Hold next step	The AI parks as soon as the current step finishes. The state goes Pause requested → Held. A held run spends no AI budget. The strip shows who is holding it and for how long.
Release AI	The run continues from where it parked.
Stop this step	The AI aborts what it is running against the target right now and moves on to its next step. The run keeps going. A confirmation dialog asks before stopping.

Things to know:

• Hold expires. A hold nobody releases is lifted automatically after a few hours (4 hours by default). The strip shows "resumes on its own in …" while the hold is armed.
• Hold does not prevent the current step. If the AI is about to write to the target, Hold will not stop that write — the strip warns "Hold parks the run after this step, so it will not prevent this write." Use Stop this step instead.
• Stop needs a step in flight. While the run is held, there is nothing to stop.
• One control at a time. Commands are sent to the agent and confirmed by it. While one is waiting for confirmation, the other controls are unavailable ("Unavailable until Hold reaches the agent."). A Stop may overtake a queued Hold.
• Honest pending state. The strip shows when a command was sent, by whom, and how long it has waited. Delivery keeps retrying. If a command does not reach the agent, the strip shows "Hold did not reach the agent." with Send again — nothing changed on the run.

Alerts in the room

The control strip raises an alert when:

• The AI is about to write to the target — decide whether to let the step finish or Stop this step.
• The AI is stuck — send a /directive in the session to unblock it.
• New Critical finding — click Review it to open it in the review queue.

When the run ends, the strip summarizes the outcome ("Run finished at …") and tells you how many findings still need a decision, with Review remaining findings.

The session composer

One input at the bottom of the Session panel handles both questions and orders. A visible switch, Send plain text as, decides where plain text goes:

Mode	Plain text goes to	Cost
Ask the assistant	The run's assistant, which answers questions about the run in the thread. The AI keeps going.	Costs run budget — answered by the assistant
Direct the AI	The agent, as a directive. The agent reads it on its next tool call.	Free — reaches the AI on its next tool call

If you type something that reads like an order while in Ask the assistant mode, the composer warns you that it will be sent as a question and suggests switching to Direct the AI.

Slash commands

Slash commands work in either mode. Type / to open the command menu; it only lists the commands that apply to the run's current state (for example, /release appears only while the run is held).

Command	Short form	What it does
/directive <text>	/d	Tell the AI what to do, now.
/chat <text>	—	Ask the assistant a question, even in Direct the AI mode.
/hold	—	Park the run after the current step.
/release	—	Let the run continue.
/stop	/st	Abort the step in flight.

Examples:

/d focus on the /api/transfer endpoint, skip static assets
/chat which endpoints accepted the forged JWT?
/hold

Slashes in text. A leading slash is only a command when the first word is a command, so directives such as focus on /admin or /api/transfer is the priority are sent as text. To send text that starts with a literal slash, type //.

Limits. Directives and questions are limited to 2,000 characters. While a question is being answered, you can still send commands.

Directives and control actions appear in the thread with their delivery status: queued, delivered, or failed.

Reviewing findings

In assisted mode, each finding the agent reports is opened as a draft issue holding the full structured report. Drafts appear in Needs your review. Nothing counts as a vulnerability until an analyst decides.

There are two kinds of card:

Kind	Meaning	Confirm action
Draft finding	A vulnerability the agent reported.	Confirm finding
Exploit attempt	The agent's validation of an exploitation.	Confirm exploit

Deciding

Open a card to see the full evidence in the detail panel: Summary, Description, Impact, Steps to reproduce, Request, Response, and Target.

• Confirm moves the draft to Identified and counts it as a vulnerability, marked as validated by analyst.
• False positive dismisses it. The dismissal is also sent to the agent so it does not keep pursuing the same lead.
• Critical and High findings are decided from the full evidence — their confirm action is only available in the detail panel, not directly on the card.
• Ask the AI about this sends the assistant a question asking how the finding was confirmed, what the evidence proves, and what the impact would be.

Undo window. A decision is recorded after 10 seconds unless you click Undo. The countdown pauses while your pointer or keyboard focus is on the undo bar. Then click Next finding to move on.

Concurrent decisions. Several analysts can review at the same time. If someone else decided the finding first, you see "Another analyst already decided this finding."

Editing before confirming

Click Edit issue in the detail panel to adjust the draft's Title, Severity, Description, and Steps to reproduce, then Save edits before confirming. Edited cards show an Edited tag.

• When severity is calculated from the CVSS score, it cannot be changed here.
• Titles are limited to 255 characters.
• If a confirmation is rejected, your edits are kept so you can fix the field and confirm again.

Findings by origin

The Findings by origin breakdown tallies the session's findings as Awaiting validation, Validated by analyst, and Dismissed, so you can see at a glance how much still needs review.

After the run ends

No new findings arrive after the run ends, but the queue stays open: the strip shows how many findings still need your decision. Confirmed findings follow the normal vulnerability management workflow, like results from any other source.

White-box findings: "Reachability unproven"

When the artifact uses a repository (Use repository), the agent also audits the source code and then tries to prove each lead live against the target.

• A source-code lead that the agent proves live is reported as a normal finding, with the white-box evidence (proof of concept and supporting documents) attached.
• A lead the agent could not prove live is still reported, labeled Reachability unproven: "Found by whitebox (source-code) analysis; its live reachability was not proven. It may exist but was not confirmed exploitable."

The label is shown on the vulnerability page. It tells reviewers the finding is real in code but not yet confirmed exploitable — it is not hidden, and it is not dressed up as confirmed.

Notifications

The pings selected on the Assisted setup card are delivered through the platform's notification channels (email and, when configured for your company, Slack). Each notification links straight to the war room.

Ping	Sent when
the run starts	The execution starts running.
a Critical is found	The agent reports a Critical finding.
the AI is about to write to the target	The agent signals it is about to perform a write action against the target.
the AI gets stuck	The agent signals it cannot make progress.
every phase change	The run moves to a new phase.

note

"About to write" and "stuck" are signaled by the agent itself. They are a best-effort warning, not a guarantee that every write is announced. Use Guardrails and Specific scope to restrict what the agent is allowed to do.

Permissions and access

Who	Can do
Users with AI Pentest create/update permissions	Create and edit assisted artifacts.
Users with AI Pentest run permission	Start and cancel runs.
Participants (lead analyst + others who can join) with update permission	Enter the war room, use controls, send directives, ask the assistant, and decide findings.
Participants with read permission only	Enter the war room and watch. Controls and decisions are read-only.
Users who are not participants	Cannot open the war room.

If your access to the company or the artifact is revoked while you are in the room, updates stop and the strip shows "Your access to this room was revoked."

Credits

An assisted run uses the same cost model as an autonomous run: credits are debited at the start of the execution based on Assessment size and Test depth, and refunded if the run fails or is cancelled. See Credits and Debit history.

Questions to the assistant are answered from the run's existing budget; they do not debit extra credits. Directives and controls are free.

Live updates

The war room updates in real time. If the live connection drops, the strip says so and the room keeps refreshing every 30 seconds until it reconnects:

• "Live updates lost for … Reconnecting; the room refreshes every 30 seconds meanwhile."
• "Live updates are unavailable. The room refreshes every 30 seconds."

Nothing is lost while disconnected: commands and decisions are stored by the platform, and the feed catches up when the connection returns.

Troubleshooting / FAQ

"This execution has no war room. Only assisted pentests have one." The execution belongs to a Single or Recurring artifact. Only assisted artifacts have a war room.

"Could not open the war room…" You are not a participant of that artifact, or the connection failed. Ask the artifact owner to add you to Others who can join, then try again.

The controls are greyed out. Either a command is still waiting for the agent to confirm it ("A command is waiting for agent confirmation."), the run has ended, or you have read-only permission.

I clicked Hold but the AI keeps going. Hold parks the run after the current step. The state shows Pause requested until the step finishes, then Held. To interrupt the current step, use Stop this step.

The run resumed by itself. A hold that nobody releases expires after a few hours (4 hours by default), and the run resumes.

"The finding is still being synchronized. Try again shortly." The draft issue is still being created on the platform. Wait a few seconds and confirm again.

"The finding was not confirmed because its severity is calculated from CVSS." Your edits are saved. Open the finding and save with the calculated severity.

"The asset is archived, so its findings can no longer change status." Unarchive the asset to decide its findings.

Can I schedule an assisted run? No. An assisted run is always started by someone who intends to watch it. Use a Recurring artifact for scheduled coverage.

Can I change an existing Single or Recurring artifact to Assisted? No — the type is locked after creation. Create a new assisted artifact.

Support

Should you have any questions or require assistance while configuring or running AI Pentest Assisted, feel free to contact our dedicated support team.