10 ways your data can leak — without any hacking at all
None of these leaks requires a hacker: they all start with "it should be fine" and often end with "can we delete this before management sees it?". The problem isn't the AI — it's the human who hands it the keys to the castle and is surprised it visits every room. Here are the 10 blind spots and the reflex that neutralizes them.
1. Screenshot
Risk — You've invested in security; the "Print Screen" key, on the other hand, costs nothing to get around.
Reflex — Mask sensitive data on screen and track exports.
2. Copy-paste
Risk — Your CISO's biggest competitor is Ctrl+C.
Reflex — Spell out in black and white what should never end up in an AI.
3. Third-party connector / plugin
Risk — A tool plugged in via an unknown URL can access far more than expected — and know others you don't see.
Reflex — Inventory connectors, install only verified sources, on a strict need-to-have basis.
4. Forgotten history
Risk — A confidential prompt from months ago may still be alive in a history.
Reflex — Enable automatic deletion of sensitive conversations.
5. Automated email
Risk — The AI sometimes sends the right message… to the wrong person.
Reflex — Mandatory human review for any sensitive external message.
6. Meeting tool
Risk — "This meeting will be recorded" — the most ignored sentence after "did you read the terms of service?".
Reflex — Define which meetings can actually be recorded.
7. Overly permissive agent
Risk — CRM, then Drive, then SharePoint, then the ERP… you've adopted a very curious agent.
Reflex — Limit access to the strict minimum, scope by scope.
8. Shared prompt
Risk — "I anonymized the data"… then comes the line "Hello Mr. Dupont".
Reflex — Create prompt templates already cleaned of any real data.
9. Poorly configured knowledge base
Risk — When a junior profile can view the executive committee's documents, that's called an incident.
Reflex — Regularly test actual access rights, not just theoretical ones.
10. Accidental training
Risk — One click on "Accept" and your data may be feeding the model.
Reflex — Check the training settings (Settings → Privacy) before any rollout.
Handling a client's confidential data? The 6-step protocol
A concrete case: a client entrusts you with their salary grid to analyze, or their general ledger to audit. These files contain personal data (compensation, employee names, suppliers): the GDPR fully applies — and the CNIL is clear, it's your organization that is legally responsible, not "the tool" nor the staff member who did the copy-paste. Here is the protocol that lets you say yes to the engagement, with confidence.
Choose the right framework — before opening Claude. On the Team and Enterprise plans, the commercial terms apply: your data is never used to train the models, and a processing agreement (DPA) is included. On an individual Pro/Max account: Settings → Privacy, disable "Help improve Claude", and never use the 👍/👎 thumbs on these conversations — the feedback sends the entire conversation to Anthropic.
Set the legal framework with the client. You act as a processor under the GDPR: mention the use of AI in the engagement letter, rely on Anthropic's DPA (at trust.anthropic.com), inform the client in writing and enter the processing in your record. Large-scale HR processing? Impact assessment (DPIA) and DPO opinion.
Minimize: send only the useful columns. A pay-equity analysis needs neither names, nor IBANs, nor addresses. Delete the identifying columns from the file before sending — it's the most effective measure, and the simplest.
Pseudonymize the rest. Replace names and employee IDs with codes (S001, S002…) and keep the mapping table locally, on your side — never in Claude. "S014 · manager · €58k" analyzes exactly as well as "Ms. Martin". You can even ask Claude to generate the pseudonymized file… on your machine, via Cowork.
Lock down the workspace. A dedicated Project for the client (isolated instructions + files), connectors and web search disabled for the duration of the analysis, deletion of the conversation at the end of the engagement if your policy requires it. The Enterprise plan adds SSO, audit logs and a custom retention period.
Deliver cleanly. Re-personalize the results locally using your mapping table, then verify that no deliverable leaves with real data toward other tools — slides, emails, shared drive.
Take a client file, delete the identifying columns and replace names with codes before sending anything.
Expected result : A pseudonymized file, and a mapping table kept locally.