Method page · Updated October 9, 2026 · Worked examples: Instinct AI terms revised August 26, 2026; Muse AI terms updated September 8, 2026, privacy policy effective September 17, 2026 · Every clause quoted below was verified at its source — the same standard this page teaches.

How to Check if an AI App Is Safe

Updated Oct 9, 2026 · 9 min read · The checklist we run on every AI agent we cover

Personal AI agents are the first consumer software that asks for your email, your files, and your payment cards at the same time — and then acts, not just answers. The makers of Muse, Meta’s agent, put the warning in their own help center: “Your Muse may be inaccurate or take unexpected actions. Monitor its actions carefully and intervene as needed.” When the vendor’s own instructions say monitor it, a checklist beats a vibe.

This is the checklist we run at Siftwire before publishing a safety page on any agent — six checks, each answerable from the product’s own documents in under an hour. Two worked examples run throughout: Instinct AI and Muse AI, the two agents we’ve checked clause by clause. Names change; the six checks don’t.

The six checks

In the order that catches the most damage earliest. Each check: the question, where to look, and what it turned up in our two worked examples.

1 · Where does training sit on day one?

Find the AI-training clause in the privacy policy and ask what the state is before you touch anything. The question isn’t “can I opt out” — it’s “what is the default while I’m deciding.”

Worked examples. Muse’s privacy policy answers in one line: “This setting is on when you first use Muse.” Instinct’s terms run training through the general clause, “subject to your opt-out rights” — also on until you act. Both defaults: on.

2 · Does an opt-out exist, and where exactly?

A control you can’t point to is a promise, not a control. The test: can you name the screen or URL without guessing?

Worked examples. Instinct’s terms name the place: “you can opt out by visiting app.instinct.com/settings.” Muse’s privacy policy puts the switch in its “improving AI models” setting. Both locatable — which is the minimum bar; anything vaguer (“manage your choices in settings”) should worry you.

3 · Which way does the opt-out reach?

This is the sentence-level check — the one that separates skimmers from readers, because the two products answer oppositely under the same headline.

Worked examples. Muse, retroactive: “Changes to this setting also apply to previous interactions.” Instinct, forward-only: “This opt out applies on a go-forward basis… We may still use AI models previously trained, fine-tuned or improved on your Materials prior to your opting out.” Same feature name, opposite meaning — and neither model can be un-trained, so “reaches backwards” still means “stops the data tap,” not “erases the past.”

4 · Is any promise stronger than the global setting?

Collect the clause-level carve-outs — promises that hold regardless of settings. These survive even when defaults change, and they tell you where the company put its real skin in the game.

Worked examples. Instinct’s Vault: “However, for Materials that you add to our feature of the Service known as the “Vault” (or its successor name), we will only use such Materials to provide the Services to you and will not use such Materials to train AI models.” Muse’s ad wall: “Muse doesn’t share your conversations or the data in your virtual machine with Meta ad systems.”

5 · What’s the cap if it goes wrong — and where do you sue?

Two numbers, both in the terms: the liability cap and the arbitration clause. Read them as the cost-of-failure line item, because that’s what they are.

Worked examples. Instinct caps total liability at “the greater of one hundred dollars ($100.00), or the amount you paid the Company Entities for the Services, if any, in the past six (6) months giving rise to the claim.” Muse’s cap — in the all-caps original — “WILL NOT EXCEED THE GREATER OF $250 OR THE AMOUNT YOU HAVE PAID US IN THE TWELVE MONTHS PRECEDING,” and explicitly covers Muse taking “ANY UNINTENDED, ERRONEOUS, OR UNAUTHORIZED ACTION.” Both route US disputes to binding arbitration; Muse’s opt-out window is 30 days, by postal mail — a deadline that expires before most people learn it exists.

6 · Does disconnect equal delete?

Never assume. Find what the documents say happens to data after you cut the connection — then find the second, separate action you’re expected to take.

Worked examples. Instinct: “…even if you disconnect a Connected Service, we may still use the indexed Connected Service Input data unless you follow the instructions to request deletion.” Muse: “After you delete something, Muse may still “remember” information it learned from what you deleted.” Both answers are no — and both name the follow-up (a deletion request; asking Muse what it remembers and using its forget function, or reading MEMORY.md directly).

Read the terms like an auditor

The six checks only work on the version of the document you actually read. Our version of that discipline, which you can copy:

Record the revision date, every time.

Both products print theirs (“updated September 8, 2026,” “revised August 26, 2026”). Write it down next to what you found; a finding without its version is a finding about nothing.

Keep your own “last verified” date.

Every safety page we publish carries one in the banner. When the vendor revises, the dated page tells you exactly which claims went stale — and re-verification is a diff, not a reread.

Verify the original language.

Policies localize. Reading Muse’s terms in October 2026, the live page served us Chinese in our region — so the English clauses were verified against an Internet Archive snapshot of the same revision instead. If you’re quoting a clause that decides money or data, quote the language you verified, and say how.

Snapshot what you agreed to.

An archive link to the terms version you accepted costs thirty seconds. If the clause changes later, that snapshot is the difference between “I agreed to this” and “I agreed to something like this.”

Read the money and dispute sections first.

Not top to bottom — targeted. The clauses that decide your worst day (agent authority, liability, arbitration, refunds) live in predictable sections, and they’re short. The marketing paragraphs at the top can wait.

Judge the approval design

Most agents now promise to “ask before acting.” The design question is whether that promise survives contact with a busy user — and the best teaching example is Meta’s own design post on Muse, which names the failure mode:

“But we also wanted to avoid banner blindness — where people approve everything to make it go away.” “How We Designed Muse” — introducing.muse.ai, September 2026
“So the default allows for standard browsing of the web and stops for anything hard to undo.” “How We Designed Muse” — introducing.muse.ai, September 2026

Three things to look for in any approval design, in that order: does the default gate the irreversible actions (the second quote is the pattern done right); does the check live outside the AI model — Muse’s help center claims its permission and security checks “operate separately from the AI model,” meaning the agent can’t talk its own gate down; and does the terms side match the design side — Instinct’s terms put it bluntly on you: “you remain responsible for configuring any additional safeguards or restrictions on Actions.”

Then verify the design claims against the incident record, not just the design post. Muse’s first month is the case study: strong written architecture, and a serious 0-day reported by Ars Technica on September 21, 2026 (hotfixed within roughly twelve hours). Both facts are true at once — which is exactly why this page has six checks and not one.

Sort the risks by what they cost you

Money — the agent’s signature is yours.

Agents that transact act as your legal agent. Instinct’s terms: “You hereby appoint the Services as your agent to enter into agreements, commitments or transactions on your behalf… binding on you as if entered into directly by you.” Muse’s: “certain transactions may be irreversible by their nature,” with recovery capped at $250 (or $100, in Instinct’s case). The rule: never connect spending authority whose worst-case loss exceeds the cap you just read.

Identity — connectors are the blast radius.

Mail, calendar, files, and credentials in one place is a single point of failure for your online identity. The Muse 0-day demonstrated the shape: one stolen token meant “complete control over the Muse account,” per Ars Technica’s reporting. Ask what one compromised credential would reach before you connect the second app.

Data — memory persists past goodbye.

Agent memory files are plaintext-readable by design (Muse’s MEMORY.md is human-readable and editable — a feature when you audit it, a target when someone else does), training is on by default in both our examples until check ① says otherwise, and check ⑥ already told you what deletion means. Data exposure isn’t one risk; it’s the residue of every other risk.

The two worked examples

The six checks, run in full, are what our two safety pages are. Each one quotes the clauses, dates the reporting, and ranks the risks by what they cost you:

Instinct AI — the worked example with the quieter record

The check found: a $100 cap, a forward-only opt-out, a Vault carve-out stronger than the opt-out, agent authority that binds — and a documented disconnect-vs-delete gap. The full record, clause by clause: Is Instinct AI safe? →

Muse AI — the worked example with the loudest first month

The check found: a $250 cap that explicitly covers the agent’s own unintended actions, a retroactive opt-out, an ad-system wall paired with monitoring rights — and the roughest launch record in the category. The full record: Is Muse AI safe? →

Every page we’ve published on either product — definitions, setup, pricing, outage records — lives on the Instinct AI hub and the Muse AI hub. When we cover a third agent, its six checks land there too, and this page gains a third worked example.

Frequently asked questions

Is it safe to connect my email to an AI app?

It depends on two things the documents state: granularity (can it only read, or also send and delete) and the liability cap if it acts wrongly. Use a dedicated email where possible, choose read-only access if the task allows it, and read checks ①–③ before the first sync. Connecting a primary inbox with send authority and a $100 cap is a decision, not a default.

How do I know if an AI app trains on my data?

Open the privacy policy and search for “train.” Three questions: is training on by default (Muse: yes, stated; most apps: yes), where is the opt-out, and which way does it reach — forward only (Instinct) or retroactive (Muse). If the policy doesn’t state the default, treat training as on.

What is a liability cap?

The maximum the company owes you no matter what its product did — Instinct’s is $100 or six months of payments (whichever is greater); Muse’s is $250 or twelve months, explicitly covering the agent’s own unintended actions. Read it as the price of the worst day, then decide what you’re willing to connect in front of it.

Can I trust an AI company’s privacy promises?

Trust the clause, not the headline. The method: verify the clause exists in the current version of the document (dated), prefer promises stronger than the global setting (check ④), and weigh the written commitments against the dated incident record — Muse shows both can be genuinely strong at the same time.

What should I do in the first hour with any AI agent?

Five moves, in order: register with a separate email; turn training off; grant the narrowest access that covers your first task; leave approvals on for anything irreversible; and note the arbitration opt-out deadline if one exists — Muse’s is 30 days by postal mail, which means it expires while you’re still exploring.

Sources

Every quotation on this page comes from documents or reporting we verified at the source, with dates — the same standard the checklist asks of you. This page adds no new product claims; it generalizes the method behind our two safety pages.