AI Pulse · Frontier Labs Watch

Anthropic’s Claude share-link feature exposed some users’ private conversation histories to others for about 90 minutes last week. The company says it patched the bug and notified affected users, but the timing could not be worse. Anthropic has been positioning itself as the enterprise-friendly AI vendor, the one that takes safety and governance seriously. Now Microsoft and Google have an opening to say, “See? Even the careful ones slip up.”

The incident matters less for what broke and more for what it signals. Enterprise buyers already worry about what happens when employees paste proprietary data into chatbots. A bug that leaks conversations between users, even briefly, confirms that worry. It does not matter if the root cause was a caching error or a routing mistake. What matters is that private chats were not private, and the vendor found out the same way everyone else did.

For companies building AI into daily work, this is a reminder that trust is not static. You can run the most locked-down model available and still hit edge cases where something leaks. That is why data governance can’t live only in the vendor’s promises. You need visibility into what your people are sharing, where it is going, and what guardrails are actually active. A system like the Omni Command Centre lets you route tasks to the right models, log what moves through them, and catch risky patterns before they become incidents.

Anthropic will recover. The bug is fixed. But Microsoft and Google will use this story in every enterprise pitch for the next six months, and they will not be wrong to do it. If you are running AI at scale, assume something will break. Plan for what you do when it does.