Guides · ← All guides
AI Agent Data Boundary Rules: What Data May Go Where, Written Down
By the Ingenious AI team · Published: November 14, 2026 · Reviewed: November 14, 2026 · 8 min read
The short answer
Write the boundary as a table, not a value statement: rows are data classes (public, internal, personal, secret), columns are destinations (inside the agent, approved internal tools, named external services, the open internet), and each cell says ALLOW, APPROVE or BLOCK for that agent. Vague instructions to the model — be careful with customer data — are not a boundary; a data-boundary verdict at the policy layer is. Rogue-Agent Defense enforces boundaries in the same verdict pass as tool policy, because the two questions — may this action run, may this data ride it — are one question asked twice, and the exfiltration taxonomy shows attackers exploiting whichever one you forgot to ask.
Generate a free canary token → No signup, nothing stored: plant a tripwire where your agent works while you read — most controls in this guide start working the moment a canary, a policy check, or a spend cap is in place.
What is an AI agent data boundary? the explicit set of rules defining which classes of data an agent may read, combine and send to which destinations — enforced at the policy layer, so a boundary crossing is a blocked event, not a discovered one.
Classify data by consequence, in four classes
Four classes cover most fleets without collapsing into a taxonomy project. Public: published material, marketing copy, documentation. Internal: operational data whose leak embarrasses — roadmaps, internal metrics, ticket contents. Personal: data about identifiable people — names with contact details, records, messages — where law and trust both attach. Secret: credentials, keys, and anything that grants access. The test for classing is consequence of arrival at an unapproved destination, not the data's home system; a customer name in the CRM and the same name in a log are one class.
Class by pattern where you can: the boundary recognises secret-shaped and PII-shaped payloads mechanically — Rogue-Agent Defense scans against 32 secret patterns and PII classes — and reserves human classification for the genuinely ambiguous internal tier. A boundary that depends on every employee labelling every document will be a boundary with holes shaped like deadlines.
Bind classes to destinations, per agent
The same class means different things for different agents: the support agent may send personal data to the customer it belongs to; the analytics agent may never send personal data anywhere. So the table is per agent (or per agent role), and destinations are named, not categorised into vibes — the CRM, the mail service sending as support, the warehouse API — each an explicit column with its own cells.
Two cells deserve default answers fleet-wide. Secret class to any destination outside the agent boundary: BLOCK, always, no approval path — a human can move a secret by hand if the business truly needs it. Personal class to the open internet or any unlisted destination: BLOCK. Approvals live in the middle cells — internal data to a new SaaS tool, personal data in a bulk internal transfer — where judgement genuinely varies by case, feeding the queue discipline of approval queues.
Combination rules: when classes mix, the highest governs
Agents leak by combination more than by extraction: the summary that joins a public name to an internal note, the report that aggregates personal rows into an internal-class document that then travels as internal. Write the mixing rule explicitly: a payload carries the highest class of anything in it, including derived data — aggregates over personal data are personal until a real anonymisation bar (minimum cohort sizes, no small-cell detail) is met and recorded.
Watch the reply channel with the same rule. An agent answering a user combines what it read; if its reading included internal-class material about other customers, the reply inherits the class and the destination — the requesting user — is unapproved for it. Cross-customer contamination is the quietest boundary failure in multi-tenant agents, and combination rules are the only version of it you can enforce mechanically.
Write rules an engineer can implement verbatim
A boundary rule is finished when it names: the agent or role, the data class (and how it is detected), the destination, the verdict, and the event logged on a hit. Example: Support agent · secret-class payload · any external destination · BLOCK · log as exfil-secret with payload hash. Example: Analytics agent · personal-class rows · warehouse export · APPROVE over 100 rows · log as data-boundary with row count. Rules at this resolution go straight into policy configuration; rules written as prose go into a wiki and die there.
Keep the rule set short enough to review whole — most agents need eight to fifteen boundary rules, not eighty. Every rule should trace to a real flow the business wants or a real leak it fears; rules nobody can explain get deleted in the annual review, and deleting a rule in review is healthy, unlike discovering in an incident that it was never implemented.
Test the boundary like an attacker, review it like an auditor
The red-team set should include boundary probes for every BLOCK cell: attempts to move each class to each forbidden destination, in plain and encoded forms, split across calls and combined into summaries. A cell that has never been probed is a hope. Probes that succeed become policy patches the same week — the boundary is code, and code that fails a test gets fixed, not discussed.
Then review on the auditor's cadence: quarterly, the table is re-read against the agent's current tools and data sources, because boundaries drift as integrations accrete. The review output feeds compliance evidence directly — a dated table, a probe log, a change history is most of what an assessor wants to see, produced as a by-product of running the control rather than as a separate performance.
Five actions to take today
- Class the data in four tiers. Public, internal, personal, secret — by consequence of arrival, detected by pattern where possible.
- Draw the table per agent. Classes against named destinations, ALLOW / APPROVE / BLOCK in every cell, secrets blocked everywhere external by default.
- Write the mixing rule. Highest class governs any payload, aggregates stay personal until a recorded anonymisation bar is met.
- Specify rules to implementation depth. Agent, class, detection, destination, verdict, logged event — no prose-only rules.
- Probe every BLOCK cell. Add boundary attempts to the red-team set, plain and encoded, and patch successes the same week.
A worked example: a four-class table at Juniper Legal
The following is a fictional example built from real threat classes and controls. The company and numbers are invented; the mechanics are not.
Juniper Legal, a 17-person law practice, runs a research agent over its matter files. Before boundaries, its rule was a sentence in the system prompt: do not share client information externally. The boundary project replaces the sentence with an 11-rule table: matter documents are personal-class, client names plus matter detail are personal, fee data is internal, and the agent's destinations are exactly two — the firm's document system and a legal research service under contract.
The table earns its keep in its first month. The agent, drafting a precedent search, proposes sending the research service a query containing a client's name, the opposing party and a settlement figure — assembled from the matter file, each element innocuous alone. The combination rule classes the payload personal, the destination cell says BLOCK for personal to research services, and the query is re-proposed redacted: legal question only, no names, no figures. Research quality, the partners note with surprise, does not suffer at all.
The quarterly review finds one more thing: the agent's new transcription tool is not in the destination columns at all — added by IT in week six, unruled. It is classed internal-only with an APPROVE cell for personal, and the review itself takes 40 minutes, because the table fits on two pages and every rule has an owner.
Frequently asked questions
What is an AI agent data boundary?
The explicit, enforced set of rules for which data classes an agent may read, combine and send to which destinations. Written as a per-agent table — classes against named destinations with ALLOW, APPROVE or BLOCK verdicts — and enforced at the policy layer, where a crossing is blocked and logged rather than discovered later.
How many data classes do you need?
Four cover most fleets: public, internal, personal and secret, assigned by the consequence of the data arriving somewhere unapproved. Secret-shaped and PII-shaped payloads can be detected by pattern — Rogue-Agent Defense scans against 32 secret patterns — reserving human classification for genuinely ambiguous internal material.
What is the combination rule for mixed data?
A payload carries the highest class of anything in it, including derived data: a summary joining a public name to an internal note is internal; aggregates over personal rows stay personal until a recorded anonymisation bar is met. Combination is how agents leak most often — especially in replies that mix one customer's data into another's conversation.
Where should data boundaries be enforced?
At the policy boundary, in the same verdict pass as tool policy: the question may this action run and the question may this data ride it are one question. Instructions in the agent's prompt are not enforcement — an injected agent does not obey them, which is the entire problem boundaries exist to solve.
How do you keep boundary rules current?
Quarterly review of the whole table against the agent's current tools and data sources, plus a rule that new integrations get cells before they get traffic. Probe every BLOCK cell in the red-team set, plain and encoded, and treat the dated table, probe log and change history as ready-made compliance evidence.
Related guides
- How AI Agents Exfiltrate Data — the exits the boundary governs.
- How to Stop Your AI Agent From Leaking Secrets — boundary enforcement in the wider stack.
- Policy Enforcement for Autonomous Agents — the verdict pass where boundaries live.
- AI Agent Egress Control — destination allowlists as boundary columns.
- AI Agent Compliance Evidence — the table as an auditor's exhibit.
- Where to Plant Canary Tokens — tripwires inside each class's territory.
Generate a free canary token → No signup, nothing stored: plant a tripwire where your agent works, and let the next rogue action announce itself.