Summary
In one of the most striking demonstrations of AI-assisted vulnerability research to date, agents built on Moonshot AI’s Kimi K3 model reportedly discovered 19 zero-day vulnerabilities in Redis — including a working remote code execution exploit for Redis 8.8.0 — in approximately 90 minutes. Researcher Chaofan Shou shared that the Kimi K3 agents autonomously identified a heap overflow in the bundled RedisBloom TDigest module and a shared-NACK double-free in Redis Streams consumer groups, then developed weaponized exploits for each class of bug.
Redis responded by releasing seven concurrent security updates on July 23, 2026 covering versions 6.2.23, 7.2.15, 7.4.10, 8.2.8, 8.4.5, 8.6.5, and 8.8.1. Exploitation of the disclosed vulnerabilities requires an authenticated session and use of the RESTORE command; no in-the-wild exploitation has been confirmed as of July 24. Mitigation options for organizations that cannot patch immediately include revoking the RESTORE command from accounts that don’t require it and restricting untrusted network access to Redis instances.
The vulnerability class overlaps with an incomplete fix from CVE-2026-25589, suggesting these issues have roots in longstanding code paths that prior security audits missed. The sheer volume of findings — 19 bugs in under two hours — underscores a new reality: AI-powered fuzzing and code analysis is now fundamentally faster than human-led vulnerability research at scale.
Source
The Hacker News — Kimi K3 Agents Found Redis Zero-Days
CyberSecurityNews — Redis Server 0-Day Exploit
Redis Security Advisories
Commentary
The implications here extend well beyond Redis. If a single AI agent session can surface 19 zero-days in a mature, well-audited open-source project in 90 minutes, we should assume that well-resourced threat actors are applying the same capability to every piece of critical infrastructure software. The attack surface isn’t growing — it’s being re-excavated at machine speed.
There’s also a secondary story here: Kimi K3 is under active U.S. government scrutiny for alleged model distillation from Anthropic. And yet its derivatives are now producing legitimate, responsible-disclosure-level security research. The line between frontier AI model capability and offensive security tooling has effectively collapsed. Defenders need to start thinking about AI-assisted patch prioritization with the same urgency that attackers are already applying AI to finding the holes.