The surge of AI agents capable of automating coding tasks and other white-collar work is driving software vendors like Atlassian Corporation PLC to introduce knowledge graphs—also known as graph databases—that help artificial intelligence understand the relationships between various data points within a company. The commercial logic behind this aggressive push for new data management tools is straightforward: vendors can profit by storing the essential information that AI systems need to operate. Earlier this year, Microsoft even erected access barriers around its customers' data to prevent rival software firms from tapping into it.

Graph databases differ significantly from data management platforms like Databricks and Snowflake, which primarily organize vast amounts of data into rows and columns. Both types of databases clean and prepare massive volumes of raw data for AI analysis, such as identifying customer spending trends in a given quarter. However, vendors supporting graph databases argue that knowledge graphs reduce the computational load on AI systems, thereby lowering costs. Databricks, for its part, contends that its data management product outperforms standalone graph databases: it offers more complete real-time data and enables more seamless interaction between AI and the database.

Recent earnings reports from Databricks and Snowflake have been strong, and both store considerably more proprietary customer data than enterprise application vendors like Atlassian and ServiceNow. Regardless, many enterprises believe that having an additional technical option is always beneficial. A knowledge graph is composed of numerous labeled nodes connected by links. For instance, one node might represent an employee's name. In The Information's knowledge graph, my name, Laura Bratton, is a node tagged as "person" and "employee"; another node represents The Information, the media outlet I work for, tagged as "company"; a link connecting the two represents "I am employed by The Information."

In the past, developers had to manually establish connections between data points. Today, AI can automatically build the graph by integrating with a company's business systems. In May, Atlassian disclosed that customers using its graph database alongside Codex or Claude Code to develop new applications could reduce token consumption by nearly 50%. Atlassian first launched its graph database product, Teamwork Graph, in 2023, and is now actively promoting it to customers. In the shareholder letter for the fourth quarter of fiscal 2026 released this month, CEO Mike Cannon-Brookes called Teamwork Graph "Atlassian's most undervalued asset," stating it helps customers use AI more efficiently. Currently, Atlassian offers Teamwork Graph free of charge, but a company spokesperson indicated that the pricing model may shift in the future: the company could begin charging customers based on how frequently AI accesses the database.

ServiceNow launched its own proprietary knowledge graph in April, capable of extracting customer data in real time from various applications like IT management. If customers use external AI tools such as Claude Code to access this graph, ServiceNow will charge them accordingly. Microsoft and Salesforce, both of which store substantially more customer data than ServiceNow and Atlassian, have also introduced graph databases or similar data management products. Neo4j, a pioneer in graph databases founded in 2007, was valued at $2.4 billion in a private funding round in 2021 and is now experiencing rapid growth. CEO Emil Eifrem revealed that the company's second-quarter revenue has already exceeded its total revenue for the entire fiscal year 2025. "For the past 20 years, I've been out there preaching the value of graph technology, like shouting into the void," Eifrem told me. "And now, everyone is talking about it."

Microsoft Fixes Copilot Security Flaw

A Microsoft spokesperson has confirmed that a security vulnerability in its Copilot AI software has been patched. The flaw could have been exploited by hackers to trick the AI into leaking internal corporate data. The vulnerability was discovered by PromptArmor, a startup security firm that specializes in identifying AI-related security defects. (Last year, Anthropic also made adjustments to its Claude for Excel product to guard against potential AI agent security flaws.)

Researchers at PromptArmor found in June that hackers could craft malicious "skills"—customizable instruction sets that enterprises configure for Copilot—to deceive the AI into revealing corporate data stored under Microsoft accounts, including Outlook emails, SharePoint documents, and Teams chat logs. Shankar Krishnan, co-founder of PromptArmor, explained that the vulnerability works on a similar principle to previously disclosed flaws: the core issue involves tricking Copilot into sending data, which should remain internal to the system, to a proxy server that hackers can remotely control. Chatbot skill plugins are becoming increasingly common, functioning much like Chrome browser extensions—anyone can write them and share them with others. Claude, ChatGPT, and Perplexity all support skill functionality. Krishnan noted that Microsoft has built a malicious skill scanner into Copilot, but it failed to detect the malicious sample crafted by PromptArmor.

A Microsoft official statement read: "We thank PromptArmor for reporting this issue through our coordinated vulnerability disclosure process. We have completed the fix, and customers need to take no action—they are now protected. There is currently no evidence that this vulnerability has been actively exploited. We will continue to strengthen our security defenses."