Security Hub / CYA Best Practices
CYA Best Practices for AI Assistants
“Cover yourself” isn't about paranoia — it's about not handing away information you can never take back. This guide turns that instinct into a short routine you can actually follow.
Educational information only — not legal, medical, or compliance advice. Confirm requirements with qualified counsel. The HIPAA, PHI, and legal concepts below are explained in general terms and do not establish or guarantee compliance with any law or framework.
The 10-minute security checklist
If you only have ten minutes, do these. Each one removes a common, costly mistake. Work top to bottom — they're ordered roughly by how much risk they remove per minute spent.
- Turn off training/history on your AI account where offered. Many tools let you opt out of having your chats used to improve the model, or disable chat history entirely. Find that toggle in privacy/data settings first.
- Use a unique password and turn on two-factor authentication. Your AI account often holds a searchable record of everything you've asked. Protect it like email — a password manager plus an authenticator app or passkey.
- Decide what you will never paste — before you're in a hurry. Write down your “never” list now (see Section 03). The mistakes happen when you're rushed and improvising.
-
Strip identifiers before you ask.
Replace real names, account numbers, and dates of birth with placeholders like
Patient Aor555-0100. The AI rarely needs the real values to help you. - Use a work/business account for work data — not your personal login. Business/enterprise tiers usually carry stronger data-handling terms. Keep personal and work histories separate.
- Treat AI output as a draft, not an authority. Verify facts, figures, citations, and code before you rely on or forward them. Models can sound confident and still be wrong.
- Keep a light record of high-stakes uses. For anything that could matter later, note the date, tool, what you asked, and what you did with the answer (see Section 04).
Settings differ by vendor and change over time. “Opted out of training” is not the same as “deleted” or “never stored.” Read the current privacy terms of the specific tool you use.
The one-page CYA rule
If you remember nothing else, remember this. It fits on a sticky note and covers the large majority of real-world AI mistakes.
Before you paste anything into an AI tool, ask: “Would I be comfortable if this exact text showed up in a screenshot, a subpoena, or a stranger's inbox?” If the answer is anything other than a clear yes — remove, redact, or don't send it.
That single question quietly handles most of the hard cases, because it forces you to picture the worst plausible destination of your text instead of the best. Three corollaries make it practical:
- When in doubt, leave it out. The cost of redacting is a few seconds. The cost of disclosure can be permanent.
- Reversibility is the test. Anything you can't “un-send” — a name, a number, a secret — deserves a pause. Generic questions don't.
- Help, then specifics. Ask the AI to solve the shape of your problem with placeholders, then fill in the real details yourself, offline.
Data classification — what to never paste
You don't need a corporate data-governance program to make good calls. You need to recognize four buckets on sight. The first is fine to share; the other three are the ones that get people in trouble.
| Class | Examples | Rule of thumb |
|---|---|---|
| Public / Generic | General questions, public facts, non-identifying drafts, sample/placeholder data like you@example.com or 555-0100. |
Safe to share. This is what AI tools are for. |
| PII — Personal Info | Real names tied to details, addresses, phone numbers, email addresses, dates of birth, government ID numbers, account numbers. | Redact or replace with placeholders. Don't paste a real person's identifying details. |
| PHI — Health Info | A named or identifiable person's diagnoses, treatments, lab results, appointments, insurance/claim details, medical record numbers. | Highest caution. Strip every identifier, or use only an approved tool covered by a proper agreement. |
| Secrets & Privileged | API keys, passwords, tokens, private keys, internal source code you don't own, attorney–client or otherwise privileged material, trade secrets. | Never paste. Treat a leaked secret as compromised — rotate it. |
A note on PHI and HIPAA (general)
Educational information only — not legal, medical, or compliance advice. Confirm requirements with qualified counsel before handling protected health information in any AI tool.
In general terms, health-privacy rules like HIPAA protect “protected health information” (PHI): health information that can be tied to a specific person. Information that is fully and properly de-identified is generally treated differently. Two practical takeaways:
- Identifiability is the trigger. “A 58-year-old with high blood pressure” is very different from that same description attached to a real name, date of birth, and record number.
- The tool matters. Putting PHI into a consumer AI account is generally not appropriate. Where PHI must be involved, organizations typically require a specific agreement and an approved, configured service — not a personal login.
What good redaction looks like
- Replace names with roles:
Patient A,Vendor 1,Manager. - Use placeholder contacts:
you@example.com,555-0100. - Round or blur dates: “mid-2025” instead of an exact birth date.
- Ask about the method, then apply it to real data yourself.
- Paste a full document “just so it has context.”
- Leave a real key in code you share — even “temporarily.”
- Assume deleting your message removes every copy.
- Rely on the AI to redact sensitive data for you.
If a real key or password ever lands in a prompt, don't just delete the message. Assume it's exposed and rotate it. A sample key for illustration should be obviously fake, e.g. sk-ant-XXXXXXXX-EXAMPLE-ONLY.
Audit trail & record-keeping
You don't need to log every chat. You do want a lightweight record for the uses that could matter later — anything touching money, health, legal exposure, employment, or a decision someone might question. A good record answers: what did the AI do, with what, and what did you do with the result?
A copy-paste log template
Keep this in a spreadsheet or a simple text note. One row per high-stakes use.
If a result is ever challenged, a short, honest record showing that a human reviewed the output — and that no sensitive data was exposed — is far more useful than trying to reconstruct events from memory months later.
Retention basics
- Keep records as long as the decision matters — and follow any retention rules that already apply to your field.
- Store the log, not the sensitive input. The point is a trail of what happened, not a second copy of the risky data.
- Review your AI account's own history settings. Know whether the tool keeps your conversations, for how long, and how to export or delete them.
Incident & breach response basics
Mistakes happen — a secret gets pasted, sensitive text goes into the wrong tool, an answer with a real name gets forwarded. What separates a small problem from a large one is what you do in the first hour. Move calmly through these steps.
- Contain it.Stop using the affected tool for that data. If a credential was exposed, rotate it immediately — assume it's already compromised.
- Capture the facts.Note what was exposed, when, in which tool, and to whom it could be visible. Don't edit history to hide it — record it accurately.
- Assess the data class.Was it Public, PII, PHI, or a secret? The class drives everything that follows. PHI and privileged material raise the stakes sharply.
- Notify the right people — internally first.Tell whoever owns security, privacy, or compliance for your organization. If you're solo, identify who you'd be obligated to inform (a client, a partner, an affected person).
- Get qualified guidance for serious cases.If PHI, privileged information, or many people's data is involved, consult qualified counsel or your privacy officer about notification obligations. Don't guess at legal duties.
- Remediate and learn.Delete what you can, document the fix, and update your “never paste” list and team rules so the same gap can't reopen.
Educational information only — not legal, medical, or compliance advice. Breach-notification duties vary by jurisdiction, contract, and data type. Confirm your specific obligations with qualified counsel.
Deleting a message may remove your copy but not every system's copy, and silent deletion can make a small mistake look worse later. Record it, fix it, and tell the people who need to know.
Team governance
For a small team, governance isn't a binder — it's a few clear agreements everyone actually knows. Three pieces cover most of it: how people start, what's acceptable, and what to do when something's unclear.
Onboarding
- Hand new people the “never paste” list and this guide on day one.
- Show them which AI tools are approved for which kinds of data — and which logins to use.
- Make sure their AI accounts have unique passwords and 2FA before they touch real work.
Acceptable-use, in plain language
- Drafting, summarizing, brainstorming, and learning with non-sensitive inputs.
- Redacting before asking, and verifying before relying.
- Asking “is this okay to share?” out loud — no penalty for caution.
- Pasting PHI, secrets, or privileged material into any unapproved tool.
- Using personal accounts for sensitive work data.
- Treating AI output as final without a human review.
Escalation
Name one person (or a shared channel) as the place to ask “can I put this in?” and to report “I think I made a mistake.” The goal is to make raising a concern faster and safer than staying quiet. A team that escalates early rarely has a large incident.
If your acceptable-use policy is longer than a page, people won't read it. The four-bucket model, the CYA rule, and a named escalation contact are enough to start.
Decision flowchart: “Can I put this into an AI tool?”
Run any text through these four questions in order. The first “yes” to a stop question means stop. Reach the bottom and you're clear to proceed.