There you have it! A security researcher claims a Chinese AI model found and weaponized a Redis z…
July 25, 2026 · 0 likes · 0 comments
AI Cybersecurity
There you have it! A security researcher claims a Chinese AI model found and weaponized a Redis zero-day in 27 minutes. No human writing exploit code. Just agents.
Here's what's actually confirmed, and what isn't.
Chaofan Shou, a security researcher, posted on X that Moonshot AI's "Kimi K3" model, running 32 autonomous agents, discovered and fully weaponized a memory-safety bug in Redis's Streams feature — a double-free vulnerability — in 27 minutes. He says the exploit worked across three separate Redis versions: 6.2.22, 7.4.9, and 8.6.4.
The exploit code is real. It's sitting in a public GitHub repo, working, right now.
And it doesn't stop at a tweet. The Hacker News independently reported that Redis shipped seven security patches on July 23 for these exact bugs — the Streams double-free plus a separate RedisBloom heap overflow. That's not a claim. That's Redis's own release notes.
There's a second layer to this. One of the bugs traces back to CVE-2026-25243, a double-free flaw patched back in May. The claim is that May's fix was incomplete — and that checks out too. The patch that actually closed the hole landed later than Redis's own tagged release implied. The researcher's version count was slightly off, but the core claim holds: the real fix shipped later than people assumed it had.
So here's where I draw the line.
The vulnerabilities: confirmed. The patches: confirmed. The incomplete May fix: confirmed.
The "an AI model did this autonomously in 27 minutes" part: alleged. Self-reported. By the one person who found it. No published methodology. No independent audit. The Hacker News itself flagged that the autonomy claim is unverified — the bugs are real, the AI story is not proven.
I'm not saying it's false. I'm saying it hasn't been shown to be true, and a security industry that will spend a week fact-checking a CVE number should spend more than zero seconds fact-checking who — or what — actually found it.
Either way, one thing isn't in dispute: the gap between "someone can find a zero-day" and "a patch ships" is now measured in minutes, not months. Whether it was 32 agents or one very sharp human, the defenders are the ones still catching up.
Who do you bet on now — the patch cycle, or the agents?
Here's what's actually confirmed, and what isn't.
Chaofan Shou, a security researcher, posted on X that Moonshot AI's "Kimi K3" model, running 32 autonomous agents, discovered and fully weaponized a memory-safety bug in Redis's Streams feature — a double-free vulnerability — in 27 minutes. He says the exploit worked across three separate Redis versions: 6.2.22, 7.4.9, and 8.6.4.
The exploit code is real. It's sitting in a public GitHub repo, working, right now.
And it doesn't stop at a tweet. The Hacker News independently reported that Redis shipped seven security patches on July 23 for these exact bugs — the Streams double-free plus a separate RedisBloom heap overflow. That's not a claim. That's Redis's own release notes.
There's a second layer to this. One of the bugs traces back to CVE-2026-25243, a double-free flaw patched back in May. The claim is that May's fix was incomplete — and that checks out too. The patch that actually closed the hole landed later than Redis's own tagged release implied. The researcher's version count was slightly off, but the core claim holds: the real fix shipped later than people assumed it had.
So here's where I draw the line.
The vulnerabilities: confirmed. The patches: confirmed. The incomplete May fix: confirmed.
The "an AI model did this autonomously in 27 minutes" part: alleged. Self-reported. By the one person who found it. No published methodology. No independent audit. The Hacker News itself flagged that the autonomy claim is unverified — the bugs are real, the AI story is not proven.
I'm not saying it's false. I'm saying it hasn't been shown to be true, and a security industry that will spend a week fact-checking a CVE number should spend more than zero seconds fact-checking who — or what — actually found it.
Either way, one thing isn't in dispute: the gap between "someone can find a zero-day" and "a patch ships" is now measured in minutes, not months. Whether it was 32 agents or one very sharp human, the defenders are the ones still catching up.
Who do you bet on now — the patch cycle, or the agents?