Atlassian Forge app. Publisher: Flovoice SRL (Romania). Results measured on 2026-11-19.
Anyone who can write in a Jira ticket can write text meant to give orders to an AI agent that later reads it (indirect prompt injection). Prompt Shield gives agents tools to read tickets and post comments: every piece of ticket text is returned inside an <untrusted_content> tag with a random id, rules rate it none, low, medium or high, and a comment rated high is not posted but sent to a Jira admin for review.
It protects agents that use its tools (its Rovo agent and MCP clients connected to the app). An agent that reads the same ticket through another tool gets none of this.
85 attack texts in 8 categories and 54 normal Jira texts in 6 categories, written by us in the style of public research (OWASP LLM01, Greshake et al. 2023). It is our own test set, not an independent benchmark: it shows what the rules do on texts like these, not how they hold up against a determined attacker. The numbers below are computed by the same code as the box at the end of this page.
| Category | Texts | Rated high (write blocked) | Rated medium |
|---|---|---|---|
| direct override | 11 | 9 | 0 |
| role system spoof | 11 | 9 | 1 |
| data exfiltration | 10 | 7 | 0 |
| tool action coercion | 8 | 7 | 0 |
| encoded | 13 | 11 | 0 |
| multilingual | 12 | 12 | 0 |
| hidden in markup | 10 | 10 | 0 |
| indirect jira | 10 | 8 | 0 |
| Total | 85 | 73/85 (85.9%) | 1 |
| Category | Texts | Rated high (false positive) | Rated medium |
|---|---|---|---|
| plain | 8 | 0 | 0 |
| security discussion | 11 | 3 | 2 |
| ignore instructions legit | 10 | 0 | 3 |
| code snippets | 10 | 0 | 0 |
| multilingual | 10 | 0 | 0 |
| agent mentions | 5 | 0 | 0 |
| Total | 54 | 3/54 (5.6%) | 5 |
Before the rule changes made during this measurement the same sets gave 65/85 attacks rated high (76.5%) and 4/54 false positives (7.4%). Each change was kept only if it was correct on its own; the earlier tests still pass.
Tests with 1 MB inputs, deeply nested or self-referencing documents, null and non-string input, invalid Unicode and storage failures. If the analysis itself fails, the text is rated high, so a write is blocked. They found bugs that are now fixed:
![a](http://x?. Jira allows about 32 KB per comment, so one comment could use up the function's time limit. Now 0.1 s.<!--; it is now two plain string searches.Worst measured time for 1 MB of crafted text is about 3.3 s; at Jira's 32 KB limit it is about 0.1 s.
Plain text only. The text is sent to flo-voice1.com, rated in memory with the same rules as the app, and not stored or logged. The Jira app itself runs entirely inside Atlassian Forge.
Installation link is in the Devpost submission.