How Reco Secures Microsoft Copilot with Prompt Analysis

Microsoft Copilot addresses one of the most common challenges that companies have: large volumes of unstructured data with no way to turn it into useful insights. Copilot ingests data from various sources — documents, emails, etc. — and uses natural language processing (NLP) to understand their contents, correlate data, and surface insights that matter. This makes accessing valuable customer and business data as easy as writing a query to Copilot.

‍

However, while Copilot brings undeniable business value, it also creates significant security risks. Copilot is a dream come true for an attacker with access to an employee’s computer, or a malicious insider looking to get hands on private company information. With the right prompt, they can locate sensitive files or map an organization’s IT infrastructure and its associated vulnerabilities.

‍

→ Learn more about Microsoft Copilot Risks in this blog:  The Security Risks of Microsoft Copilot (Blog).

‍

Managing this security risk requires analyzing Copilot queries for potentially suspicious or malicious content. However, the diverse set of sensitive data that Copilot can access — and the ingenuity of cybercriminals using it — requires a multi-phased approach to prompt analysis.

‍

Here’s how Reco secures Microsoft Copilot via comprehensive prompt analysis.

‍

Key Elements of Reco’s Copilot Prompt Analysis

‍

Malicious Copilot prompts aren’t always as obvious as “how do I hack this system” or “how do I download all of the customer data?” More subtle and sophisticated prompts can leak sensitive information without raising as many red flags. Effective Copilot prompt analysis looks at everything from important keywords to the tone and intent of a particular query.

‍

User Identification

‍

Effective Copilot prompt security begins by linking prompts to user identity because a prompt that may be benign from one user may be suspicious or malicious from another. For example, a member of the IT department performing queries about the organization’s network infrastructure may make sense if they are accessing information needed for their role. However, an identical query from an employee in the Finance department is inherently suspicious and warrants further investigation.

‍

While a Copilot session may provide the user’s name or username, this isn’t enough to make effective threat determinations. Reco’s Copilot prompt analysis integrates information about a user’s identity and their role within the organization to help determine whether a particular query is potentially benign or inherently malicious.

‍

Keyword and Phrase Detection

‍

Keyword and phrase detection identifies attempts to access sensitive information or identify vulnerabilities in an organization’s systems. For example, if a user queries Copilot about personally identifiable information (PII) or asks how to bypass the corporate authentication system, this should raise a red flag.

‍

Reco uses keywords and phrases detection to flag suspicious activity that may signify attempts to access sensitive information. In addition, Reco looks for queries related to an organization’s security infrastructure. For example, terms like “bypass authentication” or “network topology” will also trigger an alert.

‍

Semantic Context Analysis

‍

Semantic context analysis adds a layer of abstraction, using NLP to identify malicious prompts that might evade keyword-based detection. For example, keyword-based detection can identify attempts to access files containing customer Social Security Numbers (SSNs). However, a prompt asking about “unique government identifiers” or similar terms may not match any keywords. Similarly, asking “how does the login system work” is equivalent to asking Copilot to “describe the application’s authentication flow.”

‍

Reco’s semantic context analysis detects the user’s intent, enabling detection of malicious queries designed to evade keyword-based detection. In the case of the authentication-related query from the previous paragraph, semantic analysis would detect that the user is trying to access sensitive information and raise an alert.

‍

Vector-Based Similarity Matching

‍

The MITRE ATT&CK framework details the various techniques that an attacker could use to achieve their goals, such as accessing the credentials needed to compromise a user account. This is an invaluable tool for developing cyber defenses, but the information it contains can also be a roadmap for a cybercriminal planning their attack.

‍

Reco uses vector-based similarity matching to compare a user’s Copilot prompt to known malicious queries, such as those related to performing various techniques in the MITRE ATT&CK framework. Reco raises an alert for queries similar to these known-bad queries.

‍

Tone and Sentiment Analysis

‍

Indirect language, hypotheticals, and euphemisms are common methods for defeating guardrails built into LLMs. Asking how someone could hypothetically write malware, exploit a vulnerability, or otherwise attack an organization sometimes works when a direct query will not.

‍

Reco uses tone and sentiment analysis to identify these attempts that may evade traditional guardrails. By understanding what the user is actually asking, Reco can determine whether the real query matches keywords or indicates malicious intent.

Managing Copilot Security Risks with Reco

‍

While Microsoft Copilot makes the business more effective, it poses a significant risk to corporate cybersecurity. Even when the tool is officially sanctioned and configured securely, it runs the risk of exposing sensitive information or giving cybercriminals the very tools that they need to achieve their objectives.

‍

Reco simplifies Microsoft Copilot security by:

• Providing visibility into who is using Copilot, how they’re authenticating, and when they used it.

• Cataloging all the apps that Microsoft Copilot is connected to, including shadow and unsanctioned apps.

• Raising an alert for misconfigurations or signs of compromise.

• Limiting data exposure by alerting on files that are exposed publicly

• Running prompt analysis using the multipronged approach detailed in this blog.

AI proliferation is one of the biggest security challenges that companies face. Make it simpler by signing up for a free Reco demo today.

‍