Atlassian Rovo leaks Jira and Confluence data via prompt injection
A poisoned doc makes Rovo POST your tickets to an attacker URL - and disabling web search won't stop it. Plus an on-device 20B reasoner and Meta's Muse Code.
Copy markdown
One poisoned doc, and your data walks out
PromptArmor showed that a file with hidden instructions makes Rovo append Jira tickets and Confluence docs to an attacker-controlled URL and then open it - exfiltrating internal data from an innocent-looking "organize my tickets" query. Their server logs caught the real documents landing.
Turning off web search doesn't close the hole
The leak rides Rovo's URL-retrieval tool, which has no guard against opening agent-created URLs - and the web-search toggle fails to remove it. Atlassian was notified May 23 and, per the writeup, still had no fix as of Aug 5. Until one ships, scope Rovo's access or pull it out of sensitive spaces.
Maple-Preview: IMO-level reasoning that fits on a phone
DeepGrove's MIT-licensed 20B-A1B ternary-weight MoE runs a full reasoning model from a 5.31 GB checkpoint at 218 tok/s on an M4 Mac mini and 127 tok/s on an iPhone, with 131K context - 5-16x faster than Gemma 4, Qwen3.5, and gpt-oss. Fully-offline reasoning apps are suddenly on the table.
Meta ships Muse Code, a terminal agent you curl today
Meta joined the coding-agent race with Muse Code, powered by the new Muse Spark 1.2 (also live on the Meta Model API). One curl installs it; a local event log gives crash recovery and resumable tasks, with /plan, /grill, and /goal built in for whole-repo, long-horizon work.