Superintelligence and Intellectual Property: Protect What You Know

Superintelligence and intellectual property meet at one point: the more capable AI becomes at absorbing and reproducing knowledge, the more your protection depends on where that knowledge is stored, who can read it and which contract terms govern it, and the less on how secret it feels. A company protects proprietary know-how, prompts and agent memory by classifying them as assets, keeping them in systems it controls, limiting what each agent can read, and reviewing access on a schedule. Legal protection still matters, but operational control is the part you can start on today.

Key takeaways

• Capable AI raises the value of whatever it can read, so decide deliberately what each agent is allowed to read.
• Your protectable knowledge is wider than documents: it includes prompts, agent memory, evaluation examples and tool configurations.
• Keep knowledge in stores you own and can export, so you can change models without losing what your company has learned.
• Most leaks come from ordinary paths such as personal accounts, vendor terms, broad access and old contractor permissions, not from dramatic attacks.
• Let agents use sensitive knowledge through approved tools instead of holding all of it in a prompt.

What changes for intellectual property as AI gets more capable?

Superintelligence, in the strict sense, is a hypothetical level of ability beyond the best human experts. You do not need it to arrive for this topic to matter. Models available now already read long documents, summarize them, imitate a writing style and reconstruct a method from examples. Each improvement makes that easier.

For most of business history, a company's know-how was protected by friction. It lived in the heads of experienced people and in scattered files that nobody had time to read in full. An AI agent connected to your shared drive removes much of that friction. That is exactly why it is useful, and it is also why a careless setup is costly.

Two effects follow. First, anything an agent can read becomes easy to retrieve, quote and move, by the right person or the wrong one. Second, a capable model used by someone else can reconstruct more than you expect from public traces such as your website, proposals and published support answers. So protection has to be selective: be clear about what truly differentiates you, guard that tightly, and stop worrying about material that was never secret.

For a wider view of how these shifts affect planning, see Superintelligence for Business: What Changes and What Stays.

What exactly counts as your proprietary knowledge?

Most leaders think first of documents. In an agent setup, the valuable material is spread across more places than that:

• Tacit know-how: how your best people price, diagnose, prioritize and handle exceptions, usually captured only in part.
• Procedures and playbooks: templates, checklists, standard answers and decision rules.
• Prompts and instructions: role definitions, tone rules, escalation rules and the reasoning you taught the agent.
• Agent memory: corrections, customer preferences and resolved cases that accumulate over time.
• Evaluation examples: the set of cases that defines what good work looks like in your business.
• Tool and workflow configurations: how the agent connects to your systems and what it may do there.
• Client and partner data: valuable, but also carrying obligations to other parties.

Two items on that list are routinely overlooked. Evaluation examples encode your standard of quality, which is hard for a competitor to copy because it reflects judgment built over years. Agent memory grows more valuable the longer the agent works, which makes it a real asset that deserves an owner, a backup and an access policy.

A note on law: not everything above is intellectual property in a formal sense. In many jurisdictions, trade secret protection depends on taking reasonable steps to keep information confidential. The practical steps in this article are the kind of steps that support that position, but ask qualified counsel how it applies where you operate.

Where does knowledge actually leak from an AI setup?

Leaks rarely look like a movie scene. They tend to follow ordinary paths that nobody reviewed:

• Personal accounts: staff paste client files or internal rules into consumer AI tools because the approved option is slower or missing.
• Vendor terms: some services retain inputs, use them to improve their products or pass them to subprocessors. The terms differ, so read them rather than assume.
• Broad retrieval: an agent that can read every folder may answer a question for someone who should never see the underlying material.
• Prompt injection: hostile text hidden in an email, web page or document can instruct an agent to reveal its instructions or memory.
• Logs and transcripts: monitoring data is useful, and it is also a second copy of your knowledge with its own retention and access rules.
• People transitions: contractors, departing employees and former champions who keep access long after their role has ended.
• Opaque training: knowledge baked into a model hosted by a provider may be difficult to inspect, remove or move elsewhere.

The OWASP Top 10 for LLM Applications lists prompt injection and sensitive information disclosure among the main risks, which is a useful reference when you brief your technical team. The takeaway for an owner is simple: every one of these paths has a control, and most controls are decisions rather than technology purchases.

Retrieval or training: which keeps your knowledge more portable?

There are two broad ways to make a model know your business. Retrieval keeps your knowledge in documents and memory that the agent consults when it answers. Fine-tuning adjusts the model itself so that patterns from your data are built into it. AWS publishes guidance comparing the two approaches, and the trade-offs are worth understanding before you commit.

From an ownership standpoint, retrieval has a clear advantage. Your knowledge stays in a store you control. You can see what is in it, correct it, restrict who reaches it, delete an entry and carry the whole store to a different model. Fine-tuning can make sense for a narrow style or behavior, but what it absorbs is harder to audit or extract, and it ties the result to a particular model and provider.

A sound default is this: keep facts, rules and experience in retrieval and memory you own, and treat training as an exception that needs a specific reason and written contract terms about the resulting model. As models improve, you replace the model and keep the memory. That is the idea behind a company layer that outlasts any single vendor, covered in How to Avoid AI Vendor Lock-In With Business Agents.

A step-by-step plan to protect what you know

This plan suits a mid-sized company and can begin with one agent role.

• Inventory the assets. List the knowledge categories above for the role you are considering, and name one accountable owner for each.
• Classify them in three tiers: open, internal and restricted. Restricted material does not reach any agent without an explicit decision.
• Decide where each asset lives. Keep documents, memory and prompts in accounts your company owns, with version history for the instructions.
• Read the vendor terms in writing. Ask about retention, use of inputs for improvement, subprocessors, export and deletion, and keep the answers on file.
• Scope access to the role. Apply least privilege, as in the OWASP guidance on authorization, so each agent reads only what its task needs. A recurring check is described in Quarterly AI Agent Permission Review: Steps and Evidence.
• Treat memory as a controlled record. Decide who may add or correct entries, review what accumulates, and make sure you can restore it, as explained in AI Agent Backup and Restore: What to Save and How to Test It.
• Handle the people side. Give staff an approved tool and short rules for what never leaves the company, and time-box outside access using the approach in Contractor Access to a Company AI Agent: Time-Boxed Rules.
• Test for leakage. Periodically ask the agent to reveal its instructions, to answer outside its permissions and to follow instructions planted in a document. Record what happens and fix the gaps.

The plan is intentionally small. Completing the first three steps for a single role is more valuable than a company-wide policy that nobody applies.

Illustrative scenario: a distributor's pricing playbook

This is an illustrative scenario, not a description of a real client.

A mid-sized distributor has a pricing approach that lives with two senior salespeople and one complicated spreadsheet. Leadership wants an agent to draft quotes. The quick path is to upload the spreadsheet, past quotes and margin notes into a general assistant using whichever accounts people already have. It works on the first day, and nobody can say afterward who has seen what.

The controlled path starts with classification. The margin formula is restricted. The general pricing principles and approved discount ranges are internal. The agent receives the principles and the ranges, and it asks a pricing tool for an approved price instead of holding the formula itself. Customer preferences go into agent memory, and the sales lead reviews new entries each week. Instructions sit in version control. A freelancer who helps with templates receives access that expires.

Now consider what an attacker or a careless transcript could expose. In the first setup, the full formula sits in the prompt and the files. In the second, a leaked prompt reveals principles and ranges, and the formula never leaves the tool. The agent is just as useful for drafting quotes, and the most sensitive knowledge is exposed to far fewer paths. The design principle generalizes: let agents use sensitive knowledge through narrow, approved tools instead of carrying all of it in context.

How a managed approach handles this, and where to start

Be Human designs, implements and operates AI agents whose identity, memory, channels, tools and permissions stay under the company's control. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Deciding what that role may read is part of that definition. Managed operations can then include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, so protection is maintained instead of set once and forgotten. Changes to what an agent can access go through an approval step, as described in How to Request and Approve Changes to a Managed AI Agent.

You can begin without any vendor. Choose one role, list the knowledge it would touch, classify it, and write down who owns each asset and where it is stored. Those pages will be useful whichever model or partner you work with later. Take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Does using an AI agent make my trade secrets public? Not by itself. Exposure depends on the vendor's terms, where the data is stored and who the agent can answer to. Check retention and training terms in writing, restrict access to what the role needs, and ask counsel how confidentiality obligations apply to your situation.

Who owns the prompts and memory an agent builds over time? That is set by your contracts, so it should be written down before the work starts. Make sure the agreement states that instructions, memory contents and evaluation examples belong to your company and can be exported in a usable form. Qualified counsel should confirm the wording for your jurisdiction.

Should we block employees from public AI tools? Blocking alone often pushes use out of sight, where you cannot guide it. A better approach is to provide an approved tool that is good enough for daily work, publish a short list of material that never leaves the company, and explain the reasons. Review the rules as the tools and the work change.

Is the arrival of superintelligence a reason to keep knowledge away from AI? No. Withholding everything gives up the benefit of agents and does not stop competitors from using them. Protect the small set of knowledge that truly differentiates you, keep the rest accessible to well-scoped agents, and keep your stores portable so that better models improve your results without taking your knowledge with them.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.