A good prompt hygiene standard fits on one page and answers five questions: which tools and accounts are approved, what data may go in, how outputs are verified, what is recorded, and who decides exceptions. Short rules get read. Long policies get skipped.
Why a standard helps
OWASP's mitigation list for sensitive information disclosure includes guidance on avoiding the input of sensitive information and clear policies on data retention, usage and deletion. A team standard is how that guidance becomes behaviour at the keyboard.
The five sections
1. Approved tools and accounts Name the assistants and account types allowed for work. Say that personal accounts are not for company data. Say who maintains the list.
2. Data rules Use three tiers so people can decide quickly.
| Tier | Examples | Rule |
|---|---|---|
| Never | Passwords, keys, tokens, customer personal data, unreleased vulnerability details | Do not enter, ever |
| Ask first | Proprietary code, internal architecture, incident detail | Approved account only, minimum needed, owner agrees |
| Fine | Public advisories, public documentation, generic questions | No restriction |
3. Verification State that output is a draft. Facts such as versions, CVE numbers and package names are checked against a primary source. Code is tested and scanned before merge. Severity and priority are decided by a person using your policy.
4. Records Note in the ticket or pull request when an assistant helped, and what data class went in. Keep it to one line. Decide where logs live and how long.
5. Exceptions and incidents Name a contact. Say what to do if something sensitive was pasted: stop, tell the contact, rotate any credential. Say there is no blame for reporting quickly.
Prompt habits to include
- State the role and the output format.
- Give the facts you want used and say to answer "not in the data provided" otherwise.
- Mark pasted third party text as untrusted data.
- Ask for reasoning you can check, not just a conclusion.
- Do not give an assistant tools or access you would not give a new contractor on day one.
Rolling it out
Hold a fifteen minute walkthrough with real examples from your own work. Revisit after thirty days and ask what was unclear or too strict. Rules people cannot follow get ignored quietly, and you want to know about that early. Related reading: what to put in an employee generative AI use policy.
Version and review
Put a date and owner at the top. Review it when you change tools, when a vendor changes its terms, and at least twice a year.
Where Safeguard fits
Safeguard handles scanning, policy gates and compliance evidence automation in your pipeline, which keeps routine security checks out of chat windows. Free scan: https://gold.safeguard.sh/scan.
Quick check
Should the standard ban assistants for security work?
That is a choice for your risk owner. Most teams get better results with clear data tiers and verification rules than with a ban that people work around.
How long should it be?
One page. If people cannot recall the never list from memory, it is too long.
Sources
- OWASP GenAI, LLM02:2025 Sensitive Information Disclosure: https://genai.owasp.org/llmrisk/llm022025-sensitive-information-disclosure/
Last reviewed: 19 October 2026.
