Sentinel Vault: Confluence AI App Runs on Atlassian
A deep dive into Sentinel Vault v6.5.0, a Confluence AI review app powered by the Forge LLMs API with zero external data egress.
Stock photo for illustration only, not from the actual event
- Sentinel Vault uses the Forge LLMs API to review Confluence content without API keys
- Qualifies for Runs on Atlassian, keeping all data contained within Atlassian cloud
- Production build v6.5.0 passed the forge eligibility check on October 1, 2026
- Developers stripped the webtrigger module from the production script to prevent data egress
For IT administrators and security teams, approving a new AI feature usually triggers immediate data governance concerns. Most third-party AI integrations require pasting an OpenAI or Anthropic API key into a settings screen, routing your page text directly to a secondary vendor and complicating compliance reviews.
The Runs on Atlassian badge represents a different approach. Atlassian's program requires apps to use exclusively Atlassian-hosted compute and storage, match host data residency, and allow administrators to control external data egress such as analytics and logs.
Crucially, vendors cannot apply for this badge manually. Atlassian applies it automatically to eligible apps directly from the deployed marketplace build. The Sentinel Vault listing page displayed version 6.5.0 carrying this badge as of September 30, 2026.
While the Runs on Atlassian badge guarantees where data can and cannot travel, security experts remind admins that it does not automatically evaluate how well an app uses its assigned installation permissions. Examining the underlying code remains a critical step for strict compliance officers.
This architectural boundary is made possible by the Forge LLMs API, which Atlassian made generally available on July 29, 2026. According to Atlassian, apps utilizing this specific API maintain their Runs on Atlassian eligibility because all data handling stays confined within the Atlassian cloud ecosystem.
Stock photo for illustration only, not from the actual event
Inspecting the manifest.yml configuration for Sentinel Vault reveals a lean setup declaring the sentinel-vault-llm module and the claude model family, with zero remote blocks or external permission scopes outside of Confluence scopes, read-only JSM Assets scopes, and app storage.
The definitive proof comes from running Atlassian's command-line eligibility tool on October 1, 2026, which yielded starkly different results across two builds of the exact same application:
- Production deployment (version 6.5.0) passed the check and qualified for the Runs on Atlassian program.
- Development deployment (version 8.98.0) failed because it utilized a webtrigger module capable of data egress.
The engineering team mitigated this by implementing a static web trigger in mid-September, ensuring that the configuration REST API cannot egress anything unexpressed in the manifest file.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment