Skip to main content
AI risk

Risk type: Tool Risk

Claude security

Claude’s risk depends on your plan and surface. Consumer chats feed training when Model Improvement is on. Claude in Chrome adds prompt injection risk.

Free, Pro, and Max plans have a Model Improvement setting. With it on, chats can stay up to 5 years. Team, Enterprise, and API use follow commercial terms. Anthropic does not train on those by default. Claude in Chrome lets Claude act on web pages. Anthropic calls it “still risky.” Use Manually approve mode for sensitive work.

Last updated

2026-10-30

Claude surfaces, data use, and controls

Summarized from Anthropic’s privacy center, help center, and blog. Sources checked October 30, 2026.
SurfaceAvailable toData and trainingMain control
Claude apps on Free, Pro, and MaxIndividualsTrains on chats when Model Improvement is on. Then kept up to 5 years.Turn off Model Improvement. Use Incognito chats.
Claude for Work (Team and Enterprise)OrganizationsNot used for training by default.Admins enable Claude in Chrome and set site lists.
Anthropic APIDevelopersNot used for training by default.Guardrails live in your own application.
Claude in ChromePaid plans, not FreeActs on pages with your existing logins.Manually approve mode and approved-site review.

Summarized from Anthropic’s privacy center, help center, and blog. Sources checked October 30, 2026.

Affected tools

  • Claude apps on Free, Pro, and Max
  • Claude for Work (Team and Enterprise)
  • Claude in Chrome
  • Anthropic API

How it usually happens in the browser

  • People paste documents, logs, tickets, and code into Claude.
  • Work chats land in personal accounts with training on.
  • Uploaded files carry customer data into the chat.
  • Claude in Chrome reads a page with hidden instructions.
  • The agent acts with your existing logins.
  • Users switch to Skip all approvals to save time.

What traditional defenses miss

  • Classic DLP watches email and storage, not prompt boxes.
  • Network inspection lacks tab-level context for encrypted sessions.
  • An agent’s clicks look like the user’s own clicks.
  • Approval prompts lose force when people click through them.

Mitigation checklist

  • Check Model Improvement in Privacy Settings before work use.
  • Use Incognito chats to keep a chat out of training.
  • Route work data through Team or Enterprise, not personal plans.
  • Use Manually approve mode for email, banking, and admin consoles.
  • Review approved sites and revoke unneeded access.
  • Have admins restrict Claude in Chrome to approved sites.
  • Review outputs before they reach customers or auditors.
  • Allowlist the AI extensions your team may install.

Where browser isolation fits

  • Shield runs the page in a browser that is not on your computer, destroyed when you close the tab.
  • Open untrusted links from Claude answers in Shield.
  • Isolation does not change what Claude reads or obeys.
  • Copied data still crosses the isolation boundary. Keep secrets out of prompts.

Limits

  • These are Anthropic’s published terms as of October 30, 2026.
  • Your organization’s contract can override public pages.
  • Attack rates come from Anthropic’s own tests.
  • We reviewed documentation and did not test Claude.

FAQs.

Does Claude train on my conversations?
On Free, Pro, and Max, it depends on Model Improvement. With it on, Anthropic can use chats for training. Incognito chats are never used. Feedback you send can be used either way.
Is Claude Team or Enterprise data used for training?
Not by default. Anthropic excludes commercial products from training by default. Claude for Work and the API are both covered.
Is Claude in Chrome safe for banking or email?
Anthropic says Claude in Chrome is “still risky.” Anthropic prohibits purchases, trades, and account creation in every mode. Email and web content can still carry injected instructions. Use Manually approve mode on those sites.
How often does prompt injection succeed against Claude in Chrome?
Anthropic publishes its own test results. In August 2025, 23.6% of attacks succeeded without mitigations. Mitigations cut that to 11.2%. In August 2026, Anthropic used stronger red-team attacks. Attacks reaching Opus 5 succeeded 3.8% of the time. That was before added safeguards. In those tests, no attack beat Opus 5 with classifiers. Anthropic still calls prompt injection “a moving target.”
How do I delete my Claude chats?
Delete a chat from your chat history. It leaves your history at once. Anthropic deletes back-end copies within 30 days. Deleted chats are not used for future training.

References

  1. 01
  2. 02
  3. 03
    How long do you store my data?Anthropic Privacy Center
  4. 04
  5. 05
  6. 06
  7. 07
  8. 08
  9. 09
    OWASP LLM01:2025 Prompt InjectionOWASP Gen AI Security Project

Keep exploring

Try it on the next page you do not trust

Legba is a Chrome extension with two modes. Ghost gives you a private browser route. Shield opens a page in an isolated browser off your device.

Free for 30 days. No card required.

Ghost. A private route for your browser. Shield. An isolated browser, off your device. Choose the mode. Close when finished.