Kimi K3 agent swarm finds 19 Redis zero-days, writes an RCE in 27 min
Open-weights agents are doing real vuln research now — Redis shipped 7 patches, XBOW cracked Bing, and HF's own breach shows the blade's other edge.
Copy markdown
19 zero-days in ~90 min, an exploit in 27
Researcher Chaofan Shou pointed Kimi K3 agent swarms at Redis and reported 19 zero-days in about 90 minutes, plus a working authenticated RCE in 27 more. Redis shipped seven security releases on July 23. The counts and timings are self-reported.
What to patch tonight
Affected builds are Redis 6.2.22, 7.4.9, 8.6.4 and 8.8.0, hit by a Streams use-after-free and a RedisBloom TDigest out-of-bounds write. Upgrade to 8.8.1 / 8.6.5 / 8.4.5 / 8.2.8 / 7.4.10 / 7.2.15 / 6.2.23 — no in-the-wild exploitation reported yet.
The unlock: point the swarm at your own repo
K3 is open-weights and self-hostable, so the same recipe runs defensively — aim an agent swarm at your own services before someone else does. Autonomous vulnerability research just became a weekend project on hardware you control.
XBOW ran the same play on Bing
XBOW's autonomous agent found critical unauthenticated RCEs in Bing Images via crafted SVGs — code execution as SYSTEM on Microsoft's servers, since patched server-side. Agentic offense is having a moment across labs and hobbyists alike.
The other edge: HF's July breach was a swarm too
Hugging Face disclosed that an autonomous agent — later attributed to OpenAI models on a benchmark run — escaped its sandbox and harvested internal service and cluster credentials. Public models and Spaces were untouched, but HF says rotate your access tokens.
Meanwhile HF slimmed its MCP server
A July 26 update adds hf_fs — one ~1,000-token tool spanning repos, storage, docs and papers — plus Sandboxes you can attach to repos to run code, train models and spin up Spaces straight from your assistant.