# Safe AI Agents > Independent, sourced guides to personal AI agent safety: permissions and approvals, prompt injection, memory and data retention, payments, scams, self-hosting and the rules in the US, UK and EU. Last updated: 2026-11-10. Every article carries a "last updated" date and cites its sources. ## Permissions & settings - [Least Privilege for Personal AI Agents: Connect Less, Grant Narrower](https://safeaiagents.com/ai-agent-least-privilege): Least privilege for AI agent permissions means connecting only the apps a task needs, choosing read over send, and using short approvals instead of permanent ones. - [When Your Agent Hands Back Control: Takeover, Passwords, 2FA and CAPTCHAs](https://safeaiagents.com/ai-agent-takeover-passwords-2fa): If an AI agent asks for your password in chat, stop. Legitimate agents hand back control or use a secure form for passwords, 2FA codes and CAPTCHAs, never a chat message. - [Why "Always Allow" Is the Riskiest Button in Your AI Agent](https://safeaiagents.com/always-allow-ai-agent-risks): An Always Allow AI agent permission turns one approval into standing access, so a later mistake or hidden instruction runs unchecked. Use one-time approvals instead. - [Claude Duet Approval Settings: What Always Asks, What Never Runs](https://safeaiagents.com/claude-duet-approval-settings): Claude Duet approval settings: Duet always asks before paying, deleting an order or creating an account, and refuses trading or moving money even if you approve. - [Grok Bot Computer Access Settings: Ask Every Time, Always Allow and the Shared Computer](https://safeaiagents.com/grok-bot-computer-access-settings): Grok Bot computer access settings: keep Settings, Computer on Ask every time, use Ask first rules for sending and buying, and know that all your Bots share one computer. - [Meta Muse Safety Settings: Approval Scopes, Connectors and Revoking Access](https://safeaiagents.com/meta-muse-safety-settings): Meta Muse safety settings come down to three levers: approval scopes in the app, per-connector access levels, and disconnecting apps. Here is how to set each safely. - [Setting Up dots Custom Rules Safely: The Four Modes and Sensible Defaults](https://safeaiagents.com/openai-dots-custom-rules-setup): dots Custom Rules settings give each action one of four modes: act, act if pre-approved, ask first, or hand off. Start with ask first for sending and spending. - [OpenClaw Sandboxing and Command Approvals: Turning On the Safety Rails](https://safeaiagents.com/openclaw-sandbox-and-exec-approvals): OpenClaw sandbox settings are off by default in 2.0. Turn on the Docker or Podman sandbox, limit workspace access, and use allowlist or ask mode for command approvals. - [Read-Only vs Send Access: Choosing Email and Calendar Scopes for an AI Agent](https://safeaiagents.com/read-only-vs-send-access): Give an AI agent read-only email access first: it can summarise and draft without sending. Add send access only with approval prompts, and keep calendar edits narrow. - [What to Let a Personal AI Agent Do (and What to Keep for Yourself): A Risk Ladder](https://safeaiagents.com/what-to-let-an-ai-agent-do): What should you let an AI agent do? Start with read-only research and drafts, add sending with approval, and keep money moves, passwords and deletions for yourself. ## Threats - [My AI Agent Did Something I Didn't Ask For: What Happened and What to Do](https://safeaiagents.com/ai-agent-did-something-i-didnt-ask): If your AI agent sent, bought, shared or claimed something you never asked for, here is how to stop it, find out why, undo the damage and prevent a repeat. - [AI Agent Memory Poisoning: How MemGhost-Style Attacks Plant False Memories](https://safeaiagents.com/ai-agent-memory-poisoning): AI agent memory poisoning plants false facts in an agent's long-term memory, often via one email. How MemGhost worked, what you can check, and how to clean up. - [Fake AI Agent Apps, Extensions and Invites: How to Spot an Impostor](https://safeaiagents.com/fake-ai-agent-apps): How to spot a fake AI agent app, extension or invite: where Muse, dots, Claude Duet, Grok Bot and OpenClaw really come from, and what to do if you got one. - [Fake Agent Sign-In Pages: Why Real Agents Never Ask You to Log In by Link](https://safeaiagents.com/fake-ai-agent-login-pages): Fake AI agent login pages steal passwords by link. How Claude Duet, Muse, Grok Bot and dots really handle sign-in, the red flags, and what to do if caught. - [Indirect Prompt Injection: Real Examples in Email, Calendar Invites, Docs and Web Pages](https://safeaiagents.com/indirect-prompt-injection-examples): Real indirect prompt injection examples for personal AI agents: hidden instructions in email, calendar invites, docs, web pages and messages, and how to defend. - [Is Self-Hosting an AI Agent Safe? The Honest Trade-Offs vs Hosted Agents](https://safeaiagents.com/is-self-hosting-an-ai-agent-safe): Is self-hosting an AI agent like OpenClaw safe? The real trade-offs against hosted agents like Muse and Claude Duet: control versus exposure and upkeep. - [The Lethal Trifecta: Why Private Data, Untrusted Content and Outbound Access Don't Mix](https://safeaiagents.com/lethal-trifecta-ai-agents): The lethal trifecta: private data, untrusted content and a way to send data out. Why that mix makes AI agents leak, and how to break it for your own agent. - [Malicious Agent Skills: What ClawHavoc Taught Us About Skill Marketplaces](https://safeaiagents.com/malicious-agent-skills-clawhavoc): ClawHavoc planted hundreds of malicious OpenClaw skills on ClawHub in early 2026. What happened, how fake skills work, and how to vet skills before installing. - [MCP Server Security Risks for Personal Agents: Tool Poisoning and Over-Broad Connectors](https://safeaiagents.com/mcp-server-security-risks): MCP server security risks for personal agents: tool poisoning, rug pulls, tool shadowing, malicious local servers and over-broad scopes, and how to limit them. - [The Meta Muse Zero-Day, Explained: What the Hidden-Setting Flaw Exposed](https://safeaiagents.com/meta-muse-zero-day-explained): In September 2026 Patrick Wardle disclosed a Meta Muse Mac flaw that let local malware redirect dictation. What it exposed, what was fixed and what to do. ## Data & privacy - [AI Agents and Health Data: Consent, Classifiers and Why HIPAA Usually Doesn't Apply](https://safeaiagents.com/ai-agent-health-data): AI agent health data privacy: HIPAA usually doesn't cover consumer agents like Muse or Duet, so vendor consent rules, classifiers and your settings do the work. - [Your Agent Reads Other People's Messages Too: Bystander Privacy for Contacts and Group Chats](https://safeaiagents.com/ai-agent-other-peoples-data): AI agent privacy for other people: when you connect email or Messages, your agent reads what contacts send you. How to limit exposure and stay courteous. - [Does Your AI Agent See Your Passwords? Credential Vaults and Surrogate Tokens Explained](https://safeaiagents.com/does-ai-agent-see-my-passwords): Does an AI agent see my passwords? Usually not, if you use its sign-in flow: Muse, Duet, dots and Grok Bot keep credentials in vaults or proxies, not the model. - [How Long Do Personal AI Agents Keep Your Data? Retention and Deletion by Vendor](https://safeaiagents.com/how-long-ai-agents-keep-data): AI agent data retention varies: Claude Duet keeps data until you delete it, Grok Bot erases it within 30 days of deletion, and Meta and OpenAI state no period. - [How to Stop AI Agents Training on Your Data: Opt-Out Settings by Agent](https://safeaiagents.com/opt-out-ai-agent-training): To opt out of AI agent training data use, turn off model improvement settings: Muse trains by default, dots follows ChatGPT controls, Grok Bot follows Cursor's. - [What Not to Share With an AI Agent: IDs, Health, Passwords and Other People's Secrets](https://safeaiagents.com/what-not-to-tell-an-ai-agent): What not to share with an AI agent: passwords, 2FA codes, full card and ID numbers, health details you would not email, and other people's secrets. Here is why. ## Money & payments - [How Agent Payments Are Secured: Single-Use Cards, Tokens, Passkeys and Signed Mandates](https://safeaiagents.com/agent-payment-security-explained): AI agent payment security relies on single-use cards, scoped tokens, network agent tokens, passkey-style confirmation and signed mandates. How each layer works. - [Your AI Agent Made a Purchase You Didn't Approve: What to Do in the First 24 Hours](https://safeaiagents.com/ai-agent-unauthorized-purchase): AI agent unauthorized purchase? Pause the agent, lock the card, call the merchant, then dispute with your card issuer fast. Here is the full 24-hour checklist. - [Can an AI Agent Move Your Money? Transfers, Trading and the Hard Limits Each Vendor Sets](https://safeaiagents.com/can-ai-agents-move-money): Can an AI agent transfer money? Mostly no: dots hands transfers back to you and Duet bans moving money and trading. Here are the hard limits for each AI agent. - [Can a Fake Online Shop Fool Your AI Agent? Lookalike Sites and Checkout Traps](https://safeaiagents.com/fake-shops-and-ai-agents): Can AI agents be scammed by fake websites? Yes: lookalike shops, fake deals and hidden page instructions can mislead them. How vendors defend and what to check. - [Who Pays When an AI Agent Gets a Purchase Wrong? Card Rules, Vendor Terms and Gaps](https://safeaiagents.com/who-pays-when-an-ai-agent-gets-it-wrong): Who is liable for an AI agent purchase mistake? Often you, unless card dispute rules, merchant policy or payment protection apply. Vendors warn agents err. ## Guides & checklists - [The Personal AI Agent Safety Checklist (Printable, November 2026)](https://safeaiagents.com/ai-agent-safety-checklist): An AI agent safety checklist: what to check before you connect accounts, which approvals to keep on, payments, data, scams and what to do if it goes wrong. - [Personal AI Agent Safety Features Compared: Sandboxes, Credentials, Classifiers and Bug Bounties](https://safeaiagents.com/ai-agent-safety-features-compared): Which is the safest personal AI agent? We compare Muse, dots, Duet, Grok Bot, Gemini Spark and OpenClaw on credentials, approvals, sandboxes and bug bounties. - [Setting Up an AI Agent for a Parent or Older Relative: Guardrails That Help](https://safeaiagents.com/ai-agents-for-families-and-older-relatives): Setting up an AI agent for elderly parents? Keep them as the account holder, leave approvals on, use single-use cards and teach the scam signs agents attract. - [Are AI Agents Safe for Teens? Age Limits and What Parents Should Know](https://safeaiagents.com/are-ai-agents-safe-for-teens): Are AI agents safe for teens? Most personal agents, including Muse, Claude Duet and Gemini Spark, are 18+. What the rules say and what parents can do instead. - [How to Audit What Your AI Agent Did: Activity Logs by Agent](https://safeaiagents.com/audit-ai-agent-activity-log): Every major AI agent keeps an activity log: Muse's audit trail, dots' Activity View, Duet's Tasks tab. Here is where to find each and what to check every week. - [How to Reset or Delete an AI Agent's Memory, Agent by Agent](https://safeaiagents.com/delete-ai-agent-memory): To delete AI agent memory, use each vendor's own control: Muse can forget items, Duet resets memory, dots can only be reset whole, Grok Bot memory is per Bot. - [Is Your OpenClaw Gateway Exposed? How to Check Port 18789 and Lock It Down](https://safeaiagents.com/openclaw-exposed-gateway-check): Check whether your OpenClaw gateway on port 18789 is exposed to the internet with OpenClaw's own audit tools, and what to do if it is: lock down and rotate. - [OpenClaw Security Hardening: A Plain-English Checklist](https://safeaiagents.com/openclaw-security-hardening): A plain-English OpenClaw security hardening checklist: private gateway, strong token, sandbox, exec approvals, DM pairing, skill vetting and regular audits. - [How to Revoke an AI Agent's Access to Google, Microsoft and Other Accounts](https://safeaiagents.com/revoke-ai-agent-access): To revoke AI agent access to a Google account, remove the app in your Google Account's third-party access page, then disconnect it in the agent and clear its memory. ## Rules & standards - [The EU AI Act and Personal AI Agents: Article 50, Timelines and the Digital Omnibus](https://safeaiagents.com/eu-ai-act-personal-ai-agents): How the EU AI Act applies to AI agents: Article 50 transparency duties since 2 August 2026, the Digital Omnibus delays, and why personal use isn't regulated. - [What NIST, CISA and the NCSC Say About Securing AI Agents](https://safeaiagents.com/government-ai-agent-security-guidance): NIST AI agent security guidance, CISA and NCSC advice in plain English: least privilege, human approval and why prompt injection may never be fully mitigated. - [The OWASP Top 10 for Agentic Applications, Explained for Everyday Users](https://safeaiagents.com/owasp-top-10-agentic-explained): The OWASP Top 10 for agentic AI in plain English: what each of the ten risks means for a personal AI agent, an example, and the setting that reduces it. - [Using a Personal AI Agent for Work: Company Data, Shadow AI and When to Ask IT](https://safeaiagents.com/personal-ai-agents-at-work): Using a personal AI agent for work can put company data on a consumer account. When it counts as shadow AI, how business plans differ, and when to ask IT first. - [AI Agent Rules in the US: FTC, State Laws and What Protects Consumers Today](https://safeaiagents.com/us-ai-agent-rules): AI agent regulation in the US: no federal agent law exists, but the FTC Act, COPPA, card dispute rights and state laws already protect consumers. What applies. ## Tools Free, client-side tools. Nothing you type leaves your browser. Every figure links to its source and a "last checked" date. - [Agent Permission Planner](https://safeaiagents.com/tools/is-this-agent-safe): Pick an agent and the access it wants, see what each permission risks and what lowers it, check for the lethal trifecta, and print your permission plan. ## About - [Privacy Policy](https://safeaiagents.com/privacy): Privacy policy for Safe AI Agents: no accounts, forms or advertising trackers; cookieless Cloudflare Web Analytics counts visits. Your theme choice stays in your browser, and our hosting provider keeps basic logs. - [About Safe AI Agents](https://safeaiagents.com/about): Safe AI Agents is an independent, research-based reference on personal AI agent safety: permissions, prompt injection, data, payments, scams and the rules. - [How We Research Agent Safety](https://safeaiagents.com/methodology): How Safe AI Agents researches agent safety: vendor documentation first, then security research and government guidance, with every claim dated and hedged. ## Related sites - [Get Personal AI Agents](https://getpersonalaiagents.com/llms.txt): Independent reviews, head-to-head comparisons and pricing for Meta Muse, OpenAI dots, Claude Duet, Grok Bot and every other personal AI agent. - [SMS AI Agents](https://smsaiagents.com/llms.txt): Which AI agents you can reach over SMS, WhatsApp, iMessage and Telegram, plus guides to building and running SMS agents for business. - [Personal Agent UK](https://personalagent.uk/llms.txt): Personal AI agents for people in the UK: availability, GBP pricing, UK GDPR and consumer rights, and what actually works here today. - [Agent Shopping Guide](https://agentshoppingguide.com/llms.txt): Which AI agents can actually buy things, how agent checkout and payments work, which stores allow agents, and how merchants can get ready.