Google Cloud on 19 September 2026 described Mantis agents that score each check-in as it arrives, rather than waiting on a slow one-off scan. Threat models use live repo metadata and a map of calls across packages. A nightly pass after submit is the second layer; a repair bot drafts a patch on the same change request for a person to review. False-positive figures in the post are Google’s own.
Artificial Intelligence··Morning
Before submit, not after the fact
Google Cloud wrote on 19 September 2026 that its Mantis agents score each check-in as it arrives, rather than waiting for a slow, one-off scan of the tree. The company presents this as a change in when review happens, not as an independent lab result. TechCrunch, The Verge and Reuters did not carry their own page of this blog post. The card therefore stays on Google Cloud’s own announcement. The date is the blog stamp, not a second newsroom’s confirmation. The bounded search found no second newsroom; the card stays on this announcement.[1]
Live metadata, not a static memo
The same post says threat models use live repo metadata and a map of calls across packages, not a static memo. That is Google’s description of its own graph. It does not come with an outside measurement of coverage. A reader should treat the call-map claim as the vendor’s architecture note. Eigen Radar is not certifying that the map is complete. The bounded search found no second newsroom; the card stays on this announcement.[1]
A second layer and a human
A nightly pass after submit is described as the second layer. A repair bot then leaves a suggested patch on the same change request, still for a person to read. False-positive or precision figures remain Google’s own measurements. Eigen Radar is not verifying those rates. That is not a promise the patch will merge itself. The bounded search found no second newsroom; the card stays on this announcement.[1]
