Gains:
- Ability to evaluate information according to sensitivity class (public / internal / confidential / top secret) and apply the 'ask class, match the tool' reflex
- Understanding the data storage, model training, collection and breach risks of AI and keeping the open and classified world separate
- Ability to adopt the minimum data principle, personal data protection and the discipline of protecting uncertain information as if it were a higher class.
Diplomacy is based on trust, and trust is based on secrecy. A negotiation position, an intelligence assessment, the content of a conversation with an interlocutor, or a yet-to-be-announced decision, in the wrong hands, can jeopardize a state's interests, the outcome of a negotiation, or even the security of people. Artificial intelligence brings tremendous speed to this field, but it also carries new and serious privacy risks. In this unit, you will learn how to protect classified and sensitive information in the age of artificial intelligence, which data can enter which vehicle, and the discipline of safe use.
This unit is the most critical security lesson of the module. One mistake here — pasting a confidential document into the wrong tool — is irreversible. Therefore, a reflex that needs to be memorized will be established here: Before giving any data to an AI tool, always ask: in what class is this data and is this tool approved for this class?
Information classification: basic distinction
Diplomatic information is often classified according to its sensitivity. Although it varies from institution to institution, the basic logic is this:
- Unclassified/public: Information that has already been published and is publicly accessible. Example: official press releases, public news.
- Restricted/internal: Internal information that is not publicly available but has low sensitivity.
- Confidential: Information whose disclosure could seriously harm the interests of the state. Example: negotiation position, internal evaluation.
- Top secret: Information whose disclosure would cause the most serious harm. Example: sensitive intelligence, critical security information.
The rule of thumb is simple: No classified information that is not publicly available will be entered into an AI tool that your institution has not approved for that class. A public AI service is, by definition, not certified for this class of information.
Tip: If you're unsure, treat the information like a senior. Saying "this is probably already known" is a dangerous guess. Until you are 100 percent sure that information is public, protect it as if it were confidential. Unnecessary caution is not a cost, it is an assurance.
Privacy risks of AI
AI tools pose different privacy risks than traditional software:
1. Data processing and storage: The text you write to a public AI service may be processed and stored on the servers of that service. You can't control who accesses it, how long it's kept, and how it's used.
2. Use in model training: Some services may use the data you enter to improve the model. This creates the risk that your confidential information will indirectly enter the model's "memory" and leak to others.
3. Aggregation risk: Pieces of information that seem harmless individually can form a sensitive picture when combined. Small pieces of information entered at different times can collectively give away a secret position.
4. Access and breach risk: Like any external system, AI services can be subject to a data breach. The data you provide may be exposed in that breach.
The solution to these risks is to establish two separate worlds: general tools for open source work; Only in-house, secure, approved systems that do not export data for classified work. Never mix the two.
Attention: Saying "I made it anonymous" is often not enough. Even if you delete the names, the context (date, location, event, number of parties) may be enough for an expert to redact identities and position. Real sensitive information is protected not by masking but by appropriate selection of tools and media.
Safe usage discipline
Set up a practical control flow:
- Determine the class. Is this data public, internal, confidential, or top secret?
- Pair the vehicle. Can this class be processed in this tool? If not, stop.
- Minimum data policy. Even on a workable job, don't give out more information than the job requires.
- Protect personal data. Personal data (name, contact, location) are processed only if necessary and on a legal basis; If unnecessary, it is masked or not given at all.
- Keep the record. Document which tool you used for which job, with which data; If there is a problem, it can be traced.
Tip: Establish a "privacy first" culture in your team: a new clerk or intern should learn the reflex "no confidential documents enter the public domain" from day one. Most violations stem not from malice but from not knowing how open a vehicle is.
three mini cases
Case 1 — Reflex prevented violation. An expert was about to summarize an interview note when he stopped and asked its classification: the note was a confidential conversation with an interlocutor. Switched to an in-house secure system instead of a public tool. A simple "ask class first" reflex prevented a breach.
Case 2 — Collection risk recognized. A team would input small, seemingly “innocuous” pieces of information into an external tool on different days. A security review showed that when combined, these pieces gave away the agenda of an upcoming negotiation. The application has been stopped; The work was moved to a secure environment.
Case 3 — Fake anonymity. A clerk deemed a document "anonymous" because he had deleted names and handed it over to the outside agent. However, the date, place and event description in the document made it clear who was meant. Deleting the name was not enough; The document should have been processed in a secure environment anyway.
Four copyable templates
1) Class and vehicle pre-check:
I will describe the work below. Tell me (don't do the job): (1) what sensitivity class might this data have (public / internal / secret / top secret), (2) should this class be processed in a public AI tool, (3) if not, what type of environment is required. Job/data description: [here]
2) Personal data scanning and masking:
List the personal/sensitive data (name, title, contact, location, health, ID number) that must be masked before processing in the text below. PROCESSING THE TEXT; Just tell me which areas need to be removed. I will do the masking and resend. Text: [here]
3) Collection risk assessment:
I'm considering exporting the following pieces of information separately to an external tool: [pieces]. Even though these may seem harmless individually, when COMBINED can they create a sensitive picture (e.g. a position, agenda, identity)? Assess the risk of collection and tell which items should not be given together.
4) Public availability confirmation:
I'm not sure if the following information is public: [information]. List what verification I need to do before deeming it public (is there an official release, who published it, date). If I'm unsure, remind me of the "treat as confidential" principle.
Weak prompt / Strong prompt
Weak prompt:
Summarize this interview note. (hidden content is pasted directly into the note)
This is the most dangerous mistake: a confidential document has entered a public medium without its class being questioned. The risk of leakage has occurred and cannot be reversed.
Powerful prompt (check first):
I want to process a meeting note. But first: this note is probably in the secret class. Tell me if this type of document can be processed in a public tool and what is the correct way? I AM NOT GIVING THE CONTENTS OF THE DOCUMENT TO YOU; I'm just asking for the correct process.
The difference is clear: the confidential content is not given at all, the correct process is asked first and the work is directed to the appropriate environment.
Class-to-vehicle mapping table
information class
example
Public AI
right environment
public
press release
suitable
general tool
Internal/service specific
inner note
Usually no
In-house approved
secret
negotiation position
never
Secure institution system
top secret
sensitive intelligence
never
highest security
personal data
credentials
Necessary+legal only
Masked/secured
Common mistakes
- Entering data without asking for the class. Before every job, the reflex "in which class is this?" is a must.
- Assuming "it's probably obvious". Uncertain information is protected as confidential.
- Mistaking name deletion as anonymization. Context can give away identities.
- Ignoring the risk of collection. Small parts can be delicate when assembled.
- Mixing two worlds. Open and classified jobs run in separate tools/environments.
In summary
Secrecy is the most critical limit of diplomatic AI use. Take information classification (public/internal/confidential/top secret), "ask class, match vehicle" reflex, minimum data principle, personal data protection and collection risk as the backbone of this unit. General tools for open source work; Use only in-house secure, approved systems for classified work. Never confuse the two; When in doubt, guard the information like a senior.
Application task
List 8 different pieces of data from your own (or imaginary) desk (some public, some sensitive). Place each one in a class and decide "in which tool can it be processed". Evaluate ambiguous ones with the "Class and vehicle pre-check" template. Select three pieces and examine whether they form a sensitive picture when combined with the “collection risk assessment”.
checklist
- [ ] I first determined the sensitivity class for each data.
- I have never entered data whose [ ] class is not suitable for a public vehicle.
- [ ] I guarded the information I was unsure of like a superior.
- [ ] I process personal data only when necessary and lawful, masked.
- [ ] I took the risk of collecting and the separation of two worlds.