Atlassian turns customer work into AI training data by default

August 18, 2026

Atlassian activated a new data contribution policy across its cloud products that by default collects Jira and Confluence customer metadata and selected in-app content from non-enterprise plans to train its AI models. It keeps de-identified metadata for up to seven years. That is likely to matter beyond Atlassian because a work suite is no longer only where teams store tasks and documents, but also a feedstock for the vendor's model training unless the customer is exempt or pays for enterprise terms. Microsoft makes this shift visible from the other side: the more business software leans on built-in generative AI, the more the control of ordinary workplace data becomes part of the product decision, not a legal footnote.

Wiz reported that its AI-powered attack agent found and exploited a vulnerability in Snowflake's snowflake-connector-net GitHub Actions workflow that had been introduced with assistance from GitHub Copilot Autofix. GitHub Actions is the service that runs software-building steps automatically when code changes, and Copilot Autofix is the coding assistant's repair function that proposes code fixes. This appears to show a new loop in everyday development work: one AI system helps write or repair code, another AI system finds the weakness, and the cost is felt by any company that now has to treat AI-generated code as something that needs the same review as human code. For Microsoft, whose Copilot sits inside that workflow, the point is not the single bug but that AI coding tools are now exposed to the same accountability tests as any other supplier in the chain.

Anthropic published an updated Claude model report in its Transparency Hub.