Kimi K3 agents reportedly found 19 Redis zero-days in 90 minutes. Only the patches are confirmed.

Chaofan Shou, a researcher affiliated with the group Bera Buddies, posted on X that agents built on Moonshot AI's Kimi K3 found 19 Redis zero-day vulnerabilities in roughly 90 minutes, and that a separate run produced a working remote-code-execution exploit against Redis 8.8.0 in 27 minutes. The claims, amplified across security-news outlets this week, are the latest entry in a pattern of AI models being credited with autonomous vulnerability discovery — this time against one of the most widely deployed pieces of infrastructure software on the internet.

What's independently confirmed: Redis shipped seven security releases on July 23 — versions 6.2.23, 7.2.15, 7.4.10, 8.2.8, 8.4.5, 8.6.5, and 8.8.1 — fixing two real vulnerability classes. The first is a Streams consumer-group bug: a crafted stream `RESTORE` payload can make two consumers share the same NACK, producing a use-after-free that can lead to remote code execution. The second is a set of out-of-bounds writes in the bundled RedisBloom and TDigest modules, also reachable through crafted `RESTORE` payloads. Redis's own GitHub release notes describe both fixes in that exact language, independent of anyone's claim about who found them.

What isn't confirmed is everything about the discovery itself. Redis's release notes assign no CVE or CVSS score to either fix, and say nothing about who reported them or how — the standard credit section Redis usually publishes with security advisories is absent here. The 19-vulnerability count, the 90-minute and 27-minute timings, and the framing of the work as autonomous agent behavior all trace back to Shou's own account and a proof-of-concept repository the group published on GitHub, not to Redis or any third-party auditor. As of this writing, no independent researcher has reproduced the timing claims, and neither Redis nor Moonshot has issued its own statement confirming how the vulnerabilities were actually found.