How do I stop my ChatGPT or Claude account from being hijacked?

Session hijacking starts on your device, so endpoint hygiene comes first: no pirated software, no unsigned installers, no browser extensions you haven't audited. Then reduce what a stolen session is worth — never paste API keys or secrets into chat, delete sensitive conversation history, and remove saved payment methods from accounts you don't actively bill. Finally, learn where each provider's sign-out-all-devices control lives so you can revoke in seconds rather than hours.

Will changing my password stop a session hijacker?

Not reliably on its own. Many providers keep existing sessions valid after a password change, so the attacker's stolen cookie keeps working. You need an explicit sign-out-everywhere or session-revocation action. And if the malware is still on your machine, both steps are pointless — the attacker simply captures the new session. Clean the device first, then revoke, then change credentials.

What should I never paste into an AI chat?

API keys, database connection strings, private keys, internal hostnames, customer PII, and unreleased proprietary code. Conversation history is durable and readable by anyone holding a valid session token, which makes it a credential store attackers actively target. If a secret has already been pasted, rotate it rather than deleting the message — deletion does not un-expose it.

How often should I revoke AI account sessions?

Quarterly as routine hygiene, and immediately after any of these triggers: an infostealer or malware detection on any device you use, a lost or stolen laptop or phone, leaving a job where you signed in on company hardware, or a provider notification about suspicious activity. Revocation is free and costs you one re-login, so err toward doing it.

Quick Answer

How to Secure AI Accounts From Session Hijacking 2026

Published: August 31, 2026

The Short Answer

The threat: infostealer malware copies your browser cookies and replays your already-authenticated session. MFA does not stop it — MFA ran at login, and the attacker skipped login.

The fix, in order: clean the device → revoke sessions → rotate keys → reduce what a session is worth.

The single highest-value habit: stop pasting secrets into chat windows.

Last verified: August 31, 2026.

Why The Usual Advice Fails Here

Standard account-security guidance is: strong unique password, turn on MFA, done. That advice closes the front door and ignores the window.

August 2026 made this concrete. Anthropic emailed a set of Claude users warning that infostealer malware on their own computers had captured active Claude login sessions, which attackers reused to drain usage limits. Anthropic force-signed-out affected accounts, deleted saved payment methods, and refunded unauthorized charges. No password was ever involved. No MFA prompt was ever bypassed — it was simply never reached.

Session tokens are bearer credentials. Whoever holds one is you, until it expires or is revoked.

Step 1 — Protect The Endpoint (Non-Negotiable)

Every other step is void on a compromised machine.

Install nothing from cracked/pirated sources. This remains the number-one infostealer delivery route, and “I only pirate from reputable forums” is not a security model.
Audit browser extensions ruthlessly. An extension with <all_urls> host permissions can read your cookies without any malware involved. Remove anything you haven’t used in a month.
Treat “fake CAPTCHA / paste this into Run” pages as malware. The ClickFix pattern — a page instructing you to paste a command into Windows Run or Terminal — is now a leading infostealer vector.
Keep OS and browser auto-updates on. Cookie-store hardening ships in browser updates.

Step 2 — Reduce What A Stolen Session Is Worth

Assume compromise will eventually happen. Minimise the payoff.

Never paste secrets into chat. API keys, connection strings, private keys, internal hostnames, customer data. Your conversation history is fully readable to anyone holding a session token, and it is often a faster path to production than attacking production. If you have already done this, rotate the secret — deleting the message does not un-expose it.

Remove saved payment methods from AI accounts you don’t need auto-billing on. Anthropic’s own incident response included wiping saved cards, which tells you where the monetization risk sits.

Prune conversation history. Delete threads containing sensitive material. Memory features that persist context across sessions widen this surface further — review what your assistant has actually retained.

Keep API keys out of browser-accessible places. A stolen session that reaches your provider console can mint fresh keys, which is why key rotation without session revocation is theatre.

Step 3 — Learn The Revocation Controls Now

Find and bookmark the sign-out-all-devices control for every AI account you hold, before you need it. On most providers it lives under Settings → Security or Settings → Sessions/Devices.

Test it once. Confirm it actually logs out your other browser. Some products’ “log out everywhere” only clears the current device — you want to know that on a calm Tuesday, not during an incident.

Step 4 — The Recovery Sequence (Order Matters)

If you suspect compromise, run these in exactly this order:

Clean or rebuild the machine. For a confirmed infostealer hit, rebuild is the default, not the fallback. Anything you do before this is wasted.
Revoke sessions everywhere — not just the AI account you were warned about. The same cookie sweep took your Google, GitHub, Slack, cloud console and email sessions.
Rotate every API key reachable from a browser: model provider keys, GitHub PATs, cloud credentials, webhook secrets.
Re-enroll MFA on critical accounts, ensuring no attacker-added authenticator persists.
Change passwords — last, because it is the least load-bearing step here.
Audit billing on every subscription, and check provider usage logs for the exposure window.

Step 5 — For Teams

Individual hygiene doesn’t scale. If you run AI tooling for an organisation:

Enforce SSO for AI vendors so revocation is centralised at the identity provider rather than per-product.
Set short session lifetimes in IdP policy for AI tools specifically. The convenience argument for multi-week sessions doesn’t survive contact with the threat model.
Alert on usage anomalies. Sudden token-consumption spikes are the earliest reliable signal of a hijacked seat.
Prefer service accounts with scoped keys over humans pasting personal keys into shared tooling.
Ban secrets-in-chat in policy and enforce it with DLP on the egress path where you can.

What Providers Still Owe Users

The real fix is device-bound session tokens — cryptographically binding a session to hardware so a copied cookie is worthless elsewhere. The standards exist. Consumer AI adoption as of August 2026 remains thin.

Until it ships, the burden sits with users, and “sign out of everything” remains a chore most people perform only after they’ve already been robbed. Do it before.