Atlassian Rides the Knowledge Graph Boom

The rise of AI agents that automate coding projects and other white-collar tasks has prompted software firms such as Atlassian to offer so-called knowledge graphs or graph databases, which help the AI analyze the relationships between different types of data inside an organization.

Software firms typically promote these new data management tools because they can help them make money from storing the information the AI requires to do its job. Firms like Microsoft have even put up some walls around their customers’ data so competing software firms won’t be able to access it, my colleague Kevin reported earlier this year.

Graph databases, also referred to as knowledge graphs, differ from other data management platforms like those offered by Databricks and Snowflake, which mostly organize huge amounts of data into columns and rows. While both types of databases clean up troves of data so AI can analyze it and spot trends in customer spending during a given quarter, for example, proponents of graph databases say they require the AI to do less processing, which can save on costs.

Databricks, meanwhile, says its data management products are better than standalone graph databases because they have more complete, real-time information, and the AI can communicate with its database more easily. Both Databricks and Snowflake have reported strong financial performance lately, and both store much more of a customer’s proprietary data than enterprise app firms like Atlassian or ServiceNow do.

In any case, many companies think it’s worth having some options. Knowledge graphs are a collection of labeled “nodes” that connect to one another. For example, one node could be the name of a company employee. For instance, in a graph at The Information, my name Laura Bratton would be one node labeled as “person” and “employee,” and another node could be the firm that employs me, The Information, labeled “company.” A connection between those two points would indicate that I work at The Information.

Previously, developers created the graphs by drawing connections between data points. But AI is now automating this process when it’s connected to a company’s business applications.

Atlassian said in May that customers that use its graph database alongside tools such as Codex and Claude Code to develop a new app, for instance, used nearly 50% fewer tokens than if they didn’t tap the database. Atlassian first launched its graph database, Teamwork Graph, in 2023 but is now advertising it heavily to customers.

In a shareholder letter this month, CEO Mike Cannon-Brookes called Teamwork Graph the “most underappreciated part of Atlassian” because of how it helps customers use AI more effectively.

Atlassian doesn’t currently charge customers for Teamwork Graph but a spokesperson said that could change in the future, when customers may be billed in part based on how often their AI taps the database.

ServiceNow in April launched its own version of a knowledge graph, which pulls customer data in real time from across its IT management and other applications. The company charges customers that use outside AI like Claude Code to tap the product.

Amazon and Salesforce, which typically store a lot more data for customers compared to ServiceNow and Atlassian, also offer graph databases or similar data management products.

Neo4j, which helped pioneer graph databases after its founding in 2007 and was privately worth $2.4 billion as of 2021, appears to be booming: The company’s second quarter revenue, for instance, was higher than its full-year 2025 revenue, according to CEO Emil Eifrem.

“For 20 years, I've been screaming into the void about the value of graphs,” Eifrem told me. “And now all of a sudden, everyone is talking about it.”

Microsoft Patches Copilot Security Flaw

Microsoft has patched a security flaw in its Copilot AI software that could allow hackers to trick the AI into leaking companies’ data, a Microsoft spokesperson confirmed. The flaw was found by researchers at PromptArmor, a small startup that has emerged as a watchdog for such AI vulnerabilities. (Anthropic last year tweaked its Claude for Excel product after The Information reported on PromptArmor’s research into potential security flaws.)

PromptArmor researchers in June found that hackers could create a malicious “skill”—a set of instructions for Copilot that companies can configure—and trick Copilot into spitting out data from across the company’s Microsoft accounts, including emails from Outlook, files from SharePoint, and Teams messages. Like other recently discovered Copilot security flaws, this one relied on getting Copilot to send data that was meant to stay within the Copilot application to a proxy server that hackers could remotely access, according to PromptArmor cofounder Shankar Krishnan.

Skills for chatbot assistants are increasingly popular, and they function similar to a Chrome browser plugin in that anyone can create one and post it online for others to use. Chatbots like Claude, ChatGPT, and Perplexity also support skills. Microsoft has a built-in skills scanner for Copilot meant to detect malicious skills, but it didn’t catch the malicious skill PromptArmor designed, Krishnan said.

“We appreciate PromptArmor for reporting this through coordinated vulnerability disclosure,” the Microsoft spokesperson said in a statement. “We have addressed the report, and customers are already protected and do not need to take any action. We have seen no indication that this has been used against customers, and we are continuously working to strengthen our protections.”