AgentComparison
Risk checklist · Documentation-led

AI agent privacy and approval checklist

An agent that can read your inbox, spend money or run code can also make a mistake in your name. Before you connect anything, you need answers to eight questions. This page gives the questions, explains why each matters, and shows the statements recorded for each agent, with the date and source links beside the claims.

Checklist updated 8 October 2026; individual source dates below78 source-led records; privacy coverage variesVendor statements, not audits

The eight questions

  1. What can it do without asking me? Look for named modes such as draft-only, ask-first or independent. Several vendors recommend setting boundaries yourself rather than relying on defaults.
  2. Which actions are hard to undo? Sending email, paying, deleting files and pushing code are different from drafting. Check that the agent pauses before those. Instinct's terms warn that safeguards may not prevent unintended actions, and Meta warns of unexpected actions from Muse.
  3. Whose identity does it act under? Wajo says Fo emails from its own address, not yours. Other agents act in your connected accounts. Know which one you are getting before it writes to your contacts.
  4. How does it pay? Wajo describes single-use cards, but its terms also allow purchases with saved methods and credential storage where authorised. Spending controls depend on your settings. A free agent plan does not make a purchase free; check both the transaction and cancellation cost before it commits.
  5. What is kept after you disconnect? Instinct's privacy policy and OpenAI's dots FAQ each say disconnecting does not automatically delete indexed or retained data. Ask how to delete it.
  6. Is your content used to train models? Read the scope, not just a homepage slogan. Comma excludes general-purpose model training except separately agreed training or fine-tuning. Wajo excludes third-party training on its homepage, but its terms permit de-identified or aggregated training uses. Instinct's public terms include an opt-out and separate Google Workspace API restrictions. A training exclusion is not the same as deleting retained memories.
  7. Can you see and edit its memory? OpenAI says individual dot memories cannot currently be viewed or edited. Check this for every agent that remembers you.
  8. Who else can see your tasks? Wajo says a human team steps in when AI alone cannot finish a task. Ask what that human can see and when.

What the vendors say, agent by agent

Each row comes from the agent's record. "Not established" means we did not find a statement in the sources we read. It does not mean the vendor has no policy.

AgentApproval and permission statementsPrivacy and retention statements
ChatGPT workspace agents
Source check 2026-10-01
End-user and shared authentication differ; App instructions do not grant access; Shared connections can act for their ownerNot established in the sources we read.

Source 1 Source 2

Claude Code
Source check 2026-10-01
Current official docs say Auto mode starts interactive terminal/VS Code sessions and uses a classifier, not routine human prompts. Manual asks for edits and commands; check actual mode, persistent rules and extensions. Current security source.Not established in the sources we read.

Source 1 Source 2 Source 3

Comma
Source check 2026-10-07
Vendor says computers stay read-only until more access is allowed; Task decisions wait in Needs Review; ask for draft-only work if preferredData Use Policy excludes customer content from own/third-party general-purpose model training except separately agreed training or fine-tuning. Durable memory/workflow state remains until deleted or account/workspace deletion; integration credentials revoked/deleted on disconnect subject to backups. Backups typically up to 35 days; legal/security exceptions apply. Vendor statements, not an audit.

Source 1 Source 2 Source 3 Source 4

Devin
Source check 2026-10-08
IDE takeover and interactive browser on cloud sessions; CLI Normal mode prompts for writes and shell commands, Smart never auto-approves installs, destructive or secret-touching commands, Bypass auto-approves all; admin security profiles can restrict network, MCP, git and GitHub CLI access, otherwise sessions run unrestrictedCognition says customer data may train models by default; paid plans can opt out, which it says stops training use and enables zero data retention with model providers (Teams: admin only; Enterprise: no training without express written consent). Devin-processed data is kept for the duration of the customer relationship unless otherwise specified. Vendor statements, not audited.

Source 1 Source 2 Source 3 Source 4

Fin
Source check 2026-10-01
Escalation and access controls need dedicated source verificationNot established in the sources we read.

Source 1 Source 2

Fo by Wajo
Source check 2026-10-07
Homepage describes draft-only, ask-before-acting and independent work.; Terms section 5.3 permits scoped future/repeated authorization and says spending limits can be set in account settings.; Fo advertises email from its own address and human assistance; detailed human-helper access outside Google Workspace rules remains unresolved.Homepage excludes third-party model training; terms permit de-identified or aggregated content use with training algorithms, and privacy notice describes AI training. Google Workspace API data has separate training and human-access restrictions. Purpose-based retention can include derived memories/embeddings. Terms permit secure credential storage when authorized, despite the homepage no-storage wording. Vendor statements, not an audit.

Source 1 Source 2 Source 3

Instinct
Source check 2026-10-01
Public terms describe connected-service access and actions, and warn safeguards may not prevent unintended actions. Check your actual approval and account settings.Public terms permit model training with a settings opt-out and safety-review exception; Vault materials are excluded. The privacy policy separately excludes data received directly through Google Workspace APIs from model training and improvement. Disconnecting a service does not delete indexed input; terms point to Workspace for deletion requests. Vendor policy statements, not an independent audit.

Source 1 Source 2 Source 3

LangGraph
Source check 2026-10-01
Developers implement oversight and permission boundariesNot established in the sources we read.

Source 1

Lucas
Source check 2026-10-02
Vendor says users approve connected accounts, external actions and data access; Vendor says it cannot take irreversible actions, such as sending email or creating events, without explicit confirmationLucas publishes a policy covering messages, connected apps and memory, with access/correction/deletion request rights and stated legal/security retention exceptions. It is not an independent privacy audit.

Source 1 Source 2

Manus
Source check 2026-10-01
Minimum permission and approval controls not verifiedNot established in the sources we read.

Source 1 Source 2 Source 3

Microsoft Copilot Studio
Source check 2026-10-01
Tenant, connector and governance scope need separate documentationNot established in the sources we read.

Source 1 Source 2

Muse
Source check 2026-10-01
Meta recommends defining boundaries and reviewing actionsNot established in the sources we read.

Source 1

OpenAI dots
Source check 2026-10-07
Plugin permissions are shared across dots, ChatGPT, ChatGPT Work and Codex.; Custom Rules add boundaries but cannot override core safeguards. Sensitive actions can require takeover or approval; scoped advance approval is supported for some actions.; Enterprise access is off by default; app authorization and local/cloud permissions are separate.OpenAI says content is encrypted in transit and at rest. Disconnecting stops new plugin access, not deletion of retained context. Delete the dot to delete its context; separately stored files, conversations and ChatGPT Memory remain separate. Individual dot memories cannot currently be viewed or edited. Business, Enterprise and Edu data is not used for training by default; personal plans use the Improve the model for everyone setting. Human safety review can still occur. Vendor statements, not an audit.

Source 1 Source 2

Poke
Source check 2026-10-02
Connect supported integrations explicitly. Detailed write approval and retention controls need further verification.Not established in the sources we read.

Source 1 Source 2 Source 3

Replit Agent
Source check 2026-10-01
Plan mode permits review before code or data changesNot established in the sources we read.

Source 1 Source 2

Salesforce Agentforce
Source check 2026-10-01
Data access, agent governance and configured actions need verificationNot established in the sources we read.

Source 1 Source 2

Cue by Manus
Source check 2026-10-01
Manus says each agent has its own email, phone number, wallet and computer, and that you set direction and make the final call. Transaction limits were not detailed in the source we read.Not established in the sources we read.

Source 1

DoorDash text ordering
Source check 2026-10-01
DoorDash says you confirm the order inside the text thread and it manages checkout. It says it learns your preferences and can optionally message you proactively. Spending limits were not detailed.Not established in the sources we read.

Source 1

Mindtrip Personal Travel Assistant
Source check 2026-10-01
Mindtrip says travelers can activate agents that watch for things that matter and surface updates, and can book through Mindtrip. Approval rules for bookings were not detailed.Not established in the sources we read.

Source 1

Wabi
Source check 2026-10-02
Wabi says you choose what it can access, it checks with you before important actions that need approval, and you can disconnect connections at any timeWabi publishes terms and a privacy policy at wabi.ai/terms and wabi.ai/privacy. We did not audit them.

Source 1 Source 2 Source 3 Source 4

szn
Source check 2026-10-02
theszn.ai says you step back in for decisions and szn handles the rest. It does not state which actions need your approval; check before connecting accounts or payment methods.szn publishes a privacy policy at theszn.ai/privacy. ithinkdiff (27 Sep 2026) reports its founder says personal data will not be used to train models; we have not found that in the policy text we read. We did not audit it.

Source 1 Source 2 Source 3 Source 4

Underdog
Source check 2026-10-02
Underdog says messages you send and bookings you approve go to their intended recipients or websites, and that you connect Google or Microsoft accounts yourself. Its pages we read do not set out which actions need your approval; check in the app.Underdog's privacy policy (effective 30 Sep 2026) says AI processing happens on your device, with no third-party AI service, and that the iPhone and iPad app sends nothing to Underdog. The Mac app keeps an account and an optional daily usage count. Its security page (updated 29 Sep 2026) lists a passed App Defense Alliance CASA Level 1 assessment by Leviathan Security Group and Google's review for Gmail and Calendar access. Penetration tests are listed as under way. These are vendor statements, not our audit.

Source 1 Source 2 Source 3 Source 4 Source 5 Source 6 Source 7

Case
Source check 2026-10-03
Case says credentials are saved once by a human through a one-time link, encrypted, and typed into the site's login page without the agent or API seeing the password. It says 2FA codes, captchas and approvals pause the run and ask a human by phone link. These are vendor statements; the agent can still act inside the logged-in accounts on that machine.Case says logins are kept in an encrypted vault and not shown to the model, and the machine keeps cookies and files between sessions. We found no privacy policy, terms page or independent audit. Anything saved on the machine persists, so treat it like handing an agent a logged-in computer.

Source 1 Source 2

Cursor cloud agents
Source check 2026-10-03
Cursor says cloud agents clone your repository and push changes to a separate branch for handoff, and that you can add secrets, restrict outbound domains and use hooks such as preToolUse and beforeShellExecution. Enterprise adds auto-run, browser and network controls and audit logs. These are vendor statements.Cursor says privacy mode, when enabled, guarantees code data is not used for training by Cursor or its model providers, and that team admins can enforce it on Teams. Cloud agents run with your repository, secrets and network settings, so scope them before use. Not independently audited by us.

Source 1 Source 2

Jules
Source check 2026-10-03
Jules says it fetches your repository and clones it to a cloud VM, then presents a plan and a diff and creates a pull request. It describes approving the plan and reviewing the diff before the pull request. Vendor statements only; we did not find a separate data-handling page in this check.Not checked in depth. We did not read Google's data-handling terms for Jules in this pass, so treat training and retention questions as unknown until you read them.

Source 1

Folk
Source check 2026-10-03
Folk's FAQ lists one-tap connections to Gmail, Google Calendar, Drive, Docs, Sheets, Outlook, Teams, Notion, Slack, Linear, Todoist, Strava and bank accounts. These are vendor statements; review each connection you grant.Folk's FAQ says it does not and will not train models on your conversations and does not sell them. We did not review its full privacy policy or any audit.

Source 1 Source 2 Source 3

Caddy
Source check 2026-10-03
Caddy's site says it connects to your calendars and chats and talks to your usual apps in the background. We did not find a list of exact permissions or approval controls.The FAQ says messages and data should stay private and are used to personalise and improve the experience. We did not find a detailed policy on training or retention in this check.

Source 1 Source 2

Fambot
Source check 2026-10-03
Fambot's site describes reading family communications and calendars to build plans. We did not find a permission or approval-control list; check it before connecting your inbox.Not checked in depth. We did not review Fambot's privacy policy for a handling of school emails and children's information in this pass.

Source 1 Source 2 Source 3

Iris
Source check 2026-10-03
Iris says agent mode can use email, calendar, browser and image tools. We did not find a list of per-action approval controls.Iris's privacy policy says it does not sell personal information or use message content to train foundation models, that content is sent to third-party model providers such as Google, OpenAI and Anthropic, and that conversation content is generally kept up to 365 days.

Source 1 Source 2 Source 3 Source 4 Source 5

Martin
Source check 2026-10-03
Martin's docs say it can text and call others on your behalf. We did not find a list of approval controls in this check.Not checked in depth. We did not review Martin's privacy policy in this pass.

Source 1 Source 2 Source 3

Miso
Source check 2026-10-03
Miso says it books with direct airline integrations and applies loyalty status and points. We did not find approval controls or a data-handling statement on the page we read.Not checked. We did not find a privacy statement on the miso.com page we read.

Source 1 Source 2

Ohai
Source check 2026-10-03
Ohai says it uses your data for calendar organisation, reminders and task coordination and never sells or shares it for advertising. We did not find a list of approval controls.Ohai's site says family information is used only to power features and is not sold or shared for advertising. Human assistants may review complex items. We did not read the full privacy policy.

Source 1 Source 2

Ollie
Source check 2026-10-03
Ollie's FAQ says calendar and email access is granted through standard OAuth consent screens and that it never sees or stores your Google or Microsoft password.Ollie's FAQ says your data is used only to provide the service and is not shared with other users, that models come from Anthropic, Google and OpenAI via commercial APIs whose terms say inputs are not used for training, and that Ollie uses layered security including OAuth.

Source 1 Source 2 Source 3

Orbits
Source check 2026-10-03
Orbits says it syncs Apple, Google and Outlook calendars, that proposed chat changes are shown for approval before applying, and that its AI recommendations can be approved or dismissed. Vendor statements only.Not checked in depth. We did not review Orbits' privacy policy in this pass.

Source 1 Source 2

Pally
Source check 2026-10-03
Pally says bookings and purchases are always with your approval. Its FAQ says it only accesses its own Pally folder in your Google Drive and that you can delete your data.Pally's FAQ says chats use WhatsApp's end-to-end encryption and files are stored in your own Google Drive, not Pally's. We did not read its privacy policy in full.

Source 1 Source 2 Source 3

Rene
Source check 2026-10-04
Rene says email and calendar use delegated access and that it never sees or stores your password. We did not find a list of per-action approval controls. The site shows it asking before booking a table.Rene says what you share is not used to train AI models, is never sold, and is never shared without your permission. We did not review the full privacy policy.

Source 1 Source 2 Source 3

Town
Source check 2026-10-04
Town's docs list three approval modes. 'Require approval' needs your explicit approval for every write action; 'Autonomous' runs all tool calls without approval; the default depends on each tool. Calendar changes require approval by default.Town's site says it does not share, sell or train on your data. We did not review the privacy policy or any audit report.

Source 1 Source 2 Source 3 Source 4

Lindy
Source check 2026-10-04
Lindy's FAQ says anything with outside impact waits for approval, including sending an email, updating a ticket, posting to another channel or publishing a document. Read-only lookups from approved sources do not require approval. Vendor statement.Not checked in depth. We did not review Lindy's privacy policy or security documentation in this pass.

Source 1 Source 2

Kiro
Source check 2026-10-04
We did not find a documented list of permission controls in this check; the site mentions a Trust setting on sessions.Not checked in depth. We did not review Kiro's data handling terms in this pass.

Source 1 Source 2

GitHub Copilot cloud agent
Source check 2026-10-04
GitHub says the agent works on a branch and you review the diff before a pull request is created. We did not read the full permissions and firewall controls in this pass.Not checked in depth. We did not review GitHub's data handling terms for the cloud agent here.

Source 1 Source 2

Lovable
Source check 2026-10-04
Lovable's docs say you own your code and customer data stored in Lovable. We did not review permission controls for connected services in this pass.Not checked in depth. We did not review Lovable's privacy policy or enterprise governance details here.

Source 1 Source 2

Google Antigravity
Source check 2026-10-04
We did not find a documented list of permission controls in this check.Not checked in depth. We did not review Google's data handling terms for Antigravity in this pass.

Source 1 Source 2

Hedwig
Source check 2026-10-04
Connects through the official Google APIs; the vendor says granted access is scoped to the features you use. It says drafts wait for your explicit apply and the agent halts at anything it cannot undo.Vendor FAQ says it is engineered to meet SOC 2 standards, with encryption in transit and at rest, and that its AI providers are bound by agreements that forbid using your data for their own purposes, including training. These are vendor statements; we did not review the privacy policy or terms in this pass.

Source 1

Bolt
Source check 2026-10-04
We did not find a documented list of permission controls in this check.Not checked in depth. We did not review Bolt's data handling terms in this pass; the vendor lists SSO, audit logs and compliance support only on the enterprise tier.

Source 1 Source 2

Perplexity Comet
Source check 2026-10-04
We did not find a documented list of permission controls in this check; the vendor FAQ has entries on privacy settings and safe browsing that we did not expand.Not checked in depth. We did not read Perplexity's privacy terms for Comet in this pass.

Source 1 Source 2 Source 3 Source 4

Halo
Source check 2026-10-05
The vendor says Halo reads sources already on your phone (mail, calendar, reminders) and connects to apps you choose through MCP. We did not find a published list of approval controls for the actions its automations take.The vendor says Halo has no backend servers or accounts, stores data on your iPhone and, if iCloud is on, syncs it through your own iCloud account. AI requests go from the phone to the provider you configure, using your own key or subscription, under that provider's policy. The privacy policy also says it may collect anonymous crash and usage data. The homepage describes Halo as an assistant that 'never leaves your iPhone'; the policy and security page show that provider requests and iCloud sync do leave the device, so read both before connecting sensitive accounts.

Source 1 Source 2 Source 3 Source 4

Hark
Source check 2026-10-05
Hark says Handoff can use your connected accounts, with saved addresses, preferences and history, to complete tasks such as orders and bookings. Hark's terms say its services can take actions on your behalf, and that local-access features use your existing login sessions and cookies. We did not find a published list of approval steps, spending limits or confirmation rules.Hark's terms (posted 30 June 2026) say it may use your content to improve its services and train its models unless you opt out in account settings, and that content is still used for training if you send feedback or it is flagged for safety review. The terms also say you are responsible for actions taken through your account and that third-party sites may block automated activity. We did not review the full privacy policy in depth.

Source 1 Source 2 Source 3 Source 4 Source 5 Source 6

Photon (Spectrum)
Source check 2026-10-05
Photon's docs say the Spectrum server owns the agent behaviour (routing, tools, memory, safety) and providers only handle each platform's connection, so what the agent can do is set by the code you write. We did not find a published list of data Photon retains from messages that pass through its hosted lines.Messages to and from your users pass through Photon's hosted iMessage lines unless you self-host. We did not read a privacy policy or terms in depth for this entry, and found no independent audit or published retention period for message content. Check those before routing customer conversations through it.

Source 1 Source 2 Source 3 Source 4 Source 5

Open Instinct (Maria Gorskikh)
Source check 2026-10-05
The README describes six trust tiers (owner, partner, family, friend, contact, stranger) enforced in code before any tool runs, plain-English grants, a policy guard in front of every tool and a spend policy file with ask-above amounts and a daily cap. Payments use a single-use Stripe Link card that the owner approves for each purchase, and logins, 2FA and payments go to the owner through desktop takeover or a Link approval. These are the project's own statements; we did not audit the code.Because you host it, your messages, files and keys sit in your own accounts and the third-party services it uses. The README says no keys are stored in the repository and keys live in your environment or a git-ignored secrets folder. We did not review the code or the security doc, and found no independent audit.

Source 1 Source 2 Source 3 Source 4 Source 5

OpenInstinct (Merit Systems)
Source check 2026-10-05
The project's site says purchases are posted in your thread with merchant, item and total for approval, and that email, calendar events and destructive changes also need approval. It says passwords and cards are encrypted before storage and that the browser worker receives a handle and fills the saved value, keeping it out of the model. Purchases approved through Link are paid from your wallet, per the README. These are the project's statements.Because you host it, your database, memory and encrypted vault live in your own Vercel project and its connected services. The README says the encryption keys are created on first use in your private Blob store and that deleting that store loses the vault key. We did not review the code or find an independent security audit.

Source 1 Source 2 Source 3 Source 4

OpenAI Codex
Source check 2026-10-06
OpenAI's docs say the agent runs with network access off by default, that locally it uses an OS-enforced sandbox limited to the workspace, and that an approval policy controls when it must stop and ask before acting. Admin controls, SSO and audit logs are listed for Business and Enterprise. These are vendor statements; we did not test them.The pricing page says Business data is not used for training by default and links to OpenAI's business-data page. We did not review the consumer-plan data terms for this entry, so check OpenAI's current privacy terms for the plan you use.

Source 1 Source 2 Source 3 Source 4

Mitra
Source check 2026-10-06
Mitra's site says you can set transfer rules, including whether it connects you directly or calls you first to ask, and that you can watch calls live, read the transcript and step in. TestingCatalog says permission settings govern access to connected apps and that Mitra Link lets you choose who can reach your Mitra. We did not find a full list of what Mitra may do without asking.Mitra's pricing FAQ says it uses your data to improve its AI systems unless you opt out in settings, never shares it with third parties without your explicit consent (unless required by law), and uses certified encryption. The pages do not name a certification. Calls are recorded and transcribed according to the site's call summary example. We did not read the privacy policy because the page did not load.

Source 1 Source 2 Source 3 Source 4 Source 5

How to use this

Start with the agent's record, then ask the vendor the questions where the table says "not established". Connect one low-risk account first, keep sensitive accounts and payment cards out until you have answers, and review what it did after the first week. We have not tested any agent. Each row keeps the date of its source check. The checklist was refreshed on 7 October 2026, but older rows were not all re-verified that day. A policy link or privacy statement is not proof of an audit or complete retention coverage. See the methodology and report an error. For personal agents specifically, see the permissions guide and Muse permissions and safety. For team platforms, see ChatGPT workspace agents access and approvals.

A ten-minute connection check

  1. Before connecting, read the consent screen. Write down whether access is read-only or permits sending, editing and deletion. A Google sign-in link and access to Gmail or Calendar data are different link types.
  2. Choose draft-only or ask-first where available. Name the recipient and budget boundaries. Ask what can run on a schedule without a new approval.
  3. Start with one low-risk account and review the first result. If the agent cannot explain which permission a task needs, leave the sensitive account disconnected.
  4. To stop Google account access, review the linked apps in Google Account settings and remove the relevant access. Google says this stops new access but does not delete copies the app already holds. Request deletion from the agent provider separately.
  5. Check chat history, durable memory, stored files and backups as separate data types. Ask when deletion finishes, what exceptions apply and whether pending schedules stop when you disconnect.

Google's linked-apps help, read 7 October 2026, explains the separate sign-in, account-linking and data-access controls. This is a checklist, not proof that a particular agent is safe.

Stop before connecting if these answers are missing

For worked policy comparisons, read Instinct, Muse and dots, Fo by Wajo versus Comma and Poke versus Instinct. These pages narrow the questions; none is a security audit.