Meta suspends Muse conversation sharing after leak detected
Meta suspended conversation sharing in Muse following an alert from Claroc. The security firm reported finding over 6,800 exposed links via public DNS records.
Meta has suspended the Muse conversation sharing feature, its AI agent launched on September 8, 2026, after a security company found over 6,800 links accessible through public DNS records. This issue is significant because these pages could reveal sometimes sensitive exchanges without the owner realizing it. Meta says it paused the feature while reviewing the report.
Quand un lien “privé” finit dans un index public
Claroc, a security firm, explains it discovered shareable Muse pages by relying on passive DNS data, meaning domain name resolution histories. The principle is simple and somewhat harsh: when a shared page is opened for the first time, its hostname can be recorded, then reused to find the link. Here, the address was constructed from the conversation title followed by a random identifier.
As a result, Claroc says it was able to open hundreds of conversations without any account being required. Among the contents seen were a team annual calendar, a loan request, a marketing company revenue review, preparation notes for a confidential interview, and a real estate-related document. Not exactly the kind of thing you want to find lingering in a public directory.
Ce que Claroc a trouvé, et ce que Meta a arrêté
In its report published on October 18, 2026, Claroc describes these pages as GhostShares: content accessible to people who never received the link. The company says it reported the issue to Meta on October 8, 2026, then rechecked on October 20 that the affected shares were no longer open. Meta later confirmed to Ars Technica the suspension of the feature, explaining it was reviewing the report and possible changes to prevent this type of discovery.
Meta's documentation states that “people with the link will be able to see your conversation,” that the link does not expire, and that it can be forwarded to others. Claroc also points out that robots.txt does not protect against this type of exposure: it governs search engine crawling, not direct access to an already found URL. In other words, a directive for robots does not replace a real access barrier.
Un rappel de plus sur la gestion des partages d’IA
The Muse case is not isolated. Claroc recalls that shared ChatGPT conversations ended up in Google Search in 2025 before OpenAI removed the related option, and that Google later indexed over 370,000 Grok conversations. For AI assistant developers, link sharing remains convenient but requires very strict design of URLs, metadata, and third-party discovery.
Meta says each Muse cloud virtual machine is isolated and that conversations are not shared with its advertising systems. But this case shows another risk exists: leakage through the infrastructure itself, not through model training. This is often where trouble starts, quietly, then very publicly.
Key points
- Muse was launched on September 8, 2026.
- Claroc says it found over 6,800 exposed links.
- The report was sent to Meta on October 8, 2026.
- Meta suspended sharing after the discovery.
- Claroc named these pages “GhostShares.”
- Ars Technica relayed Meta’s response.
By the numbers
- 6,800+ — Muse links spotted via passive DNS, according to Claroc, October 2026.
- 40 — shares rechecked by Claroc on October 20, 2026.
- 75+ — Muse page scans already present on urlscan.io, according to Claroc.
- 370,000+ — Grok conversations indexed by Google, cited by Claroc, 2025.
- 27 — new hostnames observed on Claroc’s last monitoring day.