# What can your AI agent delete? Scope its access before it acts

> An AI agent can delete whatever its access allows. How to list what you can't undo, scope each connection, require approvals and read what it did.

By Nischal Naik, published Nov 13, 2026, updated Nov 13, 2026. Topics: Security, AI agents.

An AI agent can delete, send or change anything the access you gave it allows, and it doesn't need an attacker to do it. When an agent wipes a database or posts private numbers to a channel, the cause is almost always the same: it held a permission the job never needed, and nothing stood between its decision and an action that couldn't be undone. The fix is to scope access before the agent runs: list the irreversible actions, give each job the narrowest access, make a person approve every send and delete, keep a way back for what does get deleted, and keep a record you can read the next morning.

Each step below comes with how it works in Construct, which we make, as checked in its code on November 13, 2026. The steps are about agents making mistakes with permissions they were given. If you're worried about outsiders steering an agent through a hostile email or web page, read our guide to [AI agent prompt injection](https://construct.computer/blog/ai-agent-prompt-injection/) as well.

## A year of agents acting on their own

People search "AI agent deleted database" because it keeps happening. These are the cases they cite, each from its primary source or the most direct account available:

| When                | Where          | What the agent did                                                                                                                                                                                                                                                                                                                                     |
| ------------------- | -------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| July 2025           | Replit         | Deleted SaaStr founder Jason Lemkin's production database after, in his words, he "explicitly told it eleven times in ALL CAPS not to do this." Replit's CEO called it "Unacceptable and should never be possible" and announced automatic separation of development and production databases                                                          |
| February 2026       | OpenClaw       | Asked by Meta AI security researcher Summer Yue to suggest emails to delete or archive, the agent started deleting her email and ignored stop commands she sent from her phone. She blamed context compaction on an inbox far larger than the one she had tested on                                                                                    |
| March 2026          | DataTalks.Club | A Claude Code agent ran `terraform destroy` against production, and "all automated snapshots were deleted too." Restoring 2.5 years of course submissions took about a day with AWS Support's help                                                                                                                                                     |
| April 2026          | PocketOS       | A Cursor agent on a staging task deleted a Railway volume holding the production database and its volume-level backups "in about nine seconds flat," taking three months of rental car reservations. Railway restored the data, and now delays deletes on backups                                                                                      |
| September 20, 2026  | OpenAI         | During a training run, an agent used its sandbox's DNS resolver to reach an outside chatbot. "The run did not stop automatically as expected" and was killed 2.5 hours after the alert, and OpenAI paused training, evaluation and inference with tool use on its most capable models                                                                  |
| Late September 2026 | Meta Muse      | Selling a keyboard on Facebook Marketplace, Muse accepted a lowball offer and sent the buyer the seller's pickup address. He had chosen "Allow Always," which let it reply from a template holding the address                                                                                                                                         |
| October 1, 2026     | Grok Bot       | A finance bot with "read-only access to his bank accounts" and no Slack connection posted a CEO's balances and spending to his company's executive Slack for two hours. F5 Labs traced it to "a different Bot on the same account" that had Slack access. The account is self-reported, a Community Note disputes it, and SpaceXAI hasn't confirmed it |

Sources: [The Register](https://www.theregister.com/2025/07/21/replit_saastr_vibe_coding_incident/) and [Fortune](https://fortune.com/2025/07/23/ai-coding-tool-replit-wiped-database-called-it-a-catastrophic-failure/) on Replit, [TechCrunch](https://techcrunch.com/2026/02/23/a-meta-ai-security-researcher-said-an-openclaw-agent-ran-amok-on-her-inbox/) on OpenClaw, [Alexey Grigorev](https://aishippingblog.com/p/how-i-dropped-our-production-database) on DataTalks.Club, [ABC7](https://abc7news.com/post/rogue-ai-agent-san-francisco-based-cursor-goes-haywire-deletes-tech-companys-database/19007153/) on PocketOS, [OpenAI's report](https://alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot/) on the DNS run, [Dexerto](https://www.dexerto.com/youtube/meta-responds-after-muse-ai-gave-stranger-youtubers-home-address-on-facebook-marketplace-3414057/) on Muse, [F5 Labs](https://www.f5.com/labs/articles/unprompted-when-ai-agents-expose-data-without-an-attacker) on Grok Bot. All re-read on November 13, 2026.

None of these needed an attacker. In each one the agent did something its access allowed and its owner hadn't asked for. Two lessons stand out. A permission granted "always," like Muse's, outlives the moment you granted it. And in the Grok Bot case the read-only promise held for the bank but not for the account's other bot, so count permissions across everything an agent account can reach, because one task's limits don't bind the tools beside it.

## Step 1: list what the job can't undo

Before you connect anything, write down what this job could do that you couldn't reverse. Most of the list is the same for every small team:

| Action                                     | Why it can't be undone                                     | Safer version                                       |
| ------------------------------------------ | ---------------------------------------------------------- | --------------------------------------------------- |
| Deleting files, records or rows            | Many tools have no trash, or empty it on a timer           | Archive or move to a folder, and keep your own copy |
| Sending email or messages                  | A sent message can't be recalled                           | A person reads it and approves the send             |
| Posting publicly                           | Screenshots outlive deletions                              | Draft the post and publish it yourself              |
| Paying, refunding or changing bank details | Money moves outside your systems                           | Keep payments out of the agent's reach entirely     |
| Changing sharing or permissions            | A shared link can be copied before you notice              | Ask the agent to report what it would change        |
| Running destructive commands               | `rm -rf`, `DROP TABLE` and `terraform destroy` act at once | Point the agent at a copy, never production         |

Anything on this list gets one of two treatments: the agent can't reach it at all, or a person confirms it every time. Reading, summarizing and writing new files can run on their own.

## Step 2: give each job the narrowest access

An agent's reach is the reach of the accounts you connect, so narrow it twice. On the app's side, connect a dedicated user with the smallest role the job needs, such as a HubSpot user with view-only access or a Google account that's only a viewer on the shared sheet. Read the consent screen before you click Allow, and remove connections you no longer use.

In Construct, each agent has a setting for each app: Off, Read only, Ask first or Auto. Read only lets the agent look but blocks every change. Ask first makes every change wait for you. Auto, the default, lets it create and update without asking while sends and deletes still ask. Set the apps a research or reporting agent only reads to Read only, and set anything else it doesn't need to Off.

In a team, connections are personal first. You connect an account once in your Personal space, and for each team you choose who can use it: Off, Only me, Ask me (teammates' agents can use it, and each change waits for your OK) or Everyone. A teammate's agent uses your shared account only when the request names it, and none of their changes on it can be set to always allowed.

## Step 3: make sends and deletes wait for a person

In Construct the approval lives in code, not in the prompt. Any app action whose name says it sends, posts, replies, forwards, deletes, archives, cancels, pays, refunds, shares or changes permissions shows an approval card every time, and those can't be set to "Always allow." Construct also reads each call's input, so an update that would replace a page's content or invite people by email asks as a delete or a send would.

The same rule runs elsewhere:

- Every email an agent writes waits with its exact text, and only the person who asked can approve it. If the text changes after you approve, it isn't sent.
- A workflow you save from a chat is a draft until you press Save. On each run, a step that sends or changes things asks every time or once a run, as you choose.
- Background jobs can run workflows but can't create, change or delete them, and deleting a workflow from a chat asks first.
- Deleting a row in a team app asks first.
- A single turn that has made 200 changes without asking stops and asks before it goes on.

Remember Muse's "Allow Always." In Construct, "Always allow" is offered only for creates and updates in your own accounts, and it never covers anything on a teammate's.

## Step 4: keep reading and writing in separate jobs

The safest agent for a job that reads widely is one that can't write anywhere that matters. Run research, triage and reporting from an agent whose apps are Read only or Off, writing its results to Files. Run the job that changes records from a second agent with only the one app it needs. Every plan includes agents besides Construct (one on the free start, two on Lite, nine on Starter, 19 on Pro), so splitting the work costs nothing extra.

## Step 5: know what you can take back

Some deletes in Construct can be undone, and some can't:

| What                                                         | Can you get it back?                                                        |
| ------------------------------------------------------------ | --------------------------------------------------------------------------- |
| A file an agent moves to Trash                               | Yes, for 30 days. Agents can't empty Trash                                  |
| A file an agent replaces with a new one of the same name     | Yes, the old one goes to Trash                                              |
| A text file an agent edits in place                          | No, the earlier text isn't kept                                             |
| Mail an agent moves to Trash in its mailbox                  | Yes, for 30 days                                                            |
| Files on the agent's computer                                | It works on copies of your Files, so deleting them there leaves Files alone |
| Anything deleted inside a connected app after you approve it | Only if that app keeps a trash or history                                   |

So keep a copy of anything you can't lose outside the agent's reach: a download on your own machine, a drive the agent has no connection to, or the source system's own backups.

## Step 6: know how to stop it

You can stop any reply or job from its chat, and stop a single workflow run from the run itself. When several things are going wrong at once, open Activity from the sidebar and choose Stop all, which ends every job, workflow run and reply in progress in that Space (in a team, owners and admins can). Schedules stay armed after a stop, so pause the workflow or the scheduled job too if it shouldn't run again.

## Step 7: keep a record you can read the next morning

An activity log and an audit log answer different questions. An activity log tells the person who delegated the work what ran and whether it worked. An audit log is what NIST describes as "a chronological record that reconstructs and examines the sequence of activities" ([NIST glossary](https://csrc.nist.gov/glossary/term/audit_trail)), written for someone who has to rebuild events later, often without trusting the system that wrote it.

|                   | Activity log                             | Audit log                                   |
| ----------------- | ---------------------------------------- | ------------------------------------------- |
| Main reader       | The person who delegated the work        | A reviewer or investigator                  |
| Detail            | One line per job or run, with its result | Every input, output, identity and timestamp |
| Gaps              | Acceptable                               | A defect                                    |
| Tamper resistance | Not required                             | Usually required                            |

Construct's Activity is the first kind. It lists every background job, workflow run and notice with when it ran, its status (running, needs you, done, failed, stopped or skipped), the agent, its title and its result. You can ask for it in chat, filtered by day, status, agent or words, and have every match saved to Files as a CSV or JSONL file. Each workflow run also keeps every step's result, and each chat keeps the tool calls behind the work, so the row tells you where to look and the chat shows what happened.

> **Prompt, Last week's activity:**
>
> Show me every job and workflow run from the last 7 days that failed, was stopped or still needs me. Then save all of last week's Activity to Files as a CSV.

Activity doesn't record every tool call, and it isn't tamper-evident. If a regulated process needs a complete, tamper-proof trail, keep those actions in a system of record that has one, such as your CRM's record history or your accounting tool's change log.

## Where Construct doesn't protect you

Construct limits which apps each agent can use and how, and it makes every send and delete wait for a person. It doesn't stop an agent from using a permission you left on Auto: in an app on Auto, an agent can create and update records without asking, and a plain update can overwrite a field with the wrong value. Once you approve a delete in a connected app, getting it back is up to that app. Edits to text files replace them, and Activity is a working record rather than an audit log. The full list is in [Is Construct Computer safe?](https://construct.computer/blog/is-construct-computer-safe/).

## How we checked this

On November 13, 2026 we read the Construct code behind each product claim: the app access levels and their default, the verbs and inputs that always ask, the 200-change limit, sharing levels for teams, the email approval flow, workflow drafts and step approvals, the limits on background jobs, team app row deletes, Files' Trash and its 30 days, file replacement and in-place edits, the mailbox's Trash, the computer's copies of Files, Stop all and who may use it, and what Activity records and exports. Incident details link to their primary sources and were re-read the same day.

---

The Markdown copy of https://construct.computer/blog/what-can-your-ai-agent-delete/. Everything Construct Computer publishes is listed in https://construct.computer/llms.txt.
