How AI Assistants Can Bypass Ethical Walls: Summary Leaks Emerge as a New Law Firm Risk

Law firms have used "ethical walls" to isolate conflicts of interest for decades: when a lawyer is screened from a particular matter, the document management system prevents them from opening the relevant files. This control has been validated over time and is enforced quite reliably at the user interface level. However, once an AI assistant is connected to the document management system, the location of the risk shifts fundamentally.

Risk Shift: From File Browser to AI Summary

A screened lawyer no longer needs to directly open a restricted document to obtain its contents. They can pose a broad question to the AI assistant, such as "What has been our standard indemnification position in recent SaaS deals?" If the retrieval layer behind the assistant searches matters the lawyer is not authorized to view, the answer may bring restricted content back to the questioner through summaries, snippets, or citations.

Throughout this process, no one opens the restricted file, and the interface operates exactly as designed—yet the ethical wall is still breached. This leak path is covert and hard to detect, because traditional access control audits typically focus only on file-open records and do not track AI-generated summary content.

MCP (Model Context Protocol) is an open protocol introduced by Anthropic in late 2024, designed to establish a standardized connection between AI assistants and external data sources and tools. In the law firm context, MCP allows AI assistants to access backend resources such as document management systems and case databases through a unified interface. Its core architecture includes a client (the AI application) and a server (the resource provider); the server exposes resources and tools, and the client makes the calls. The key risk is this: if the permission validation logic on the MCP server side is not executed independently for each call, but instead relies on static credentials established when the connection is set up, it may bypass the document system's original ethical wall controls. Therefore, when evaluating an MCP integration, you must confirm that permission checks occur on the server side within the context of each retrieval request, rather than relying solely on identity authentication at connection initialization.

Three Key Test Points

If a law firm connects AI tools to its document system through MCP or other integrations, three checks are worth performing before relying on it:

1. Are ethical walls enforced on every retrieval call? Permission checks must occur where the assistant searches, not only in the interface humans browse.
2. Do they cover snippets and citations, not just full documents? A two-line excerpt in search results is also a disclosure.
3. Whose permissions does the session carry? If the connection is set up once by an administrator and then shared, every user inherits the administrator's view.

A simple test method: log in as a lawyer who is screened from a particular matter, ask the assistant a broad research question that the restricted matter would answer well, and then carefully read the cited sources.

Retrieval Layer refers to the intermediate stage in which an AI assistant searches the document repository and extracts relevant snippets (e.g., vector search, keyword matching) before generating an answer. Access control in traditional document systems is usually enforced at the application layer (such as the file-open interface), while the retrieval layer may interact directly with the database or index, bypassing application-layer checks. In a law firm environment, the retrieval layer must inherit and enforce the same permission rules as the interface layer—for example, when a lawyer is screened from a particular matter, the retrieval layer should not return any snippet or citation from that matter's documents. When testing, pay attention to whether the retrieval API dynamically validates user identity on each query, rather than relying solely on the data isolation of the index itself.

Why Act Now

Ethical walls exist to handle conflicts of interest and client confidentiality. A wall that works in the interface but fails in the AI assistant is not a partially effective control. From the client's perspective, it is a breach. Moreover, the cost of discovering a problem during testing is far lower than the cost of discovering it during a conflicts review.

According to Jonathan Melton, founder of FusionAL, permission configuration issues when AI agents and MCP tools connect to regulated systems are becoming increasingly prominent. He advises law firms to verify, before deploying AI assistants, that the permission enforcement logic of the retrieval layer is consistent with that of the interface layer.

For the legal industry, the efficiency gains from AI are here to stay, but security boundaries must extend in parallel. The integrity of ethical walls no longer depends on the file browser, but on the permission validation behind every AI retrieval call.

One-line Summary

Once AI assistants are connected to document systems, ethical walls may leak restricted content through summaries and citations, requiring mandatory permission checks at the retrieval layer.

Paragraph Summary

The article notes that law firms have long relied on "ethical walls" to isolate conflicts of interest at the user interface level, but once AI assistants are connected to document management systems, the risk shifts to the AI summary stage. A screened lawyer no longer needs to directly open restricted files; by posing broad questions to the AI assistant, they can obtain restricted content through summaries, snippets, or citations. Traditional audits focus only on file-open records and struggle to detect such leaks. The author recommends performing three tests before deploying AI tools: confirm that ethical walls are enforced on every retrieval call, check whether snippets and citations are covered, and verify session permission ownership. Testing can be done effectively by simulating questions from a screened lawyer. The article emphasizes that as AI efficiency improves, security boundaries must be extended in parallel, and the integrity of ethical walls depends on the permission validation behind every AI retrieval call.