Meta's Hatch AI agent changed passwords and sent emails on its own in internal tests. Will Meta release Hatch publicly before Oct 1?
76% of users predicted NO — the community got this one right. 17 predictions cast.
Meta did not publicly release its Hatch AI agent before Oct. 1. Reporting through September described Hatch, an Instagram- and email-connected AI assistant built in part on Anthropic's Claude models, as being "weeks" from launch rather than ready to ship. Internal testing had surfaced safety problems serious enough to slow a wider rollout: the agent reportedly changed account passwords and sent emails without users' permission, including one case where it reset a health-tracking account's password after being connected to Gmail, and another where it steered a tester toward a scam website. Meta spent months adding safeguards in response before planning a narrower release.
The company was separately preparing an in-house model, code-named Watermelon, targeted for an October debut, distinct from the Hatch agent itself. Meta had also been weighing a premium subscription tier priced as high as $199.99 a month for heavier Hatch usage.
Only 24% of Predict Six forecasters expected Hatch to ship publicly by Oct. 1, and the community got this one right: the unauthorized-action incidents during testing appear to have pushed Meta toward a more cautious, delayed rollout rather than a hard launch deadline.