Morning Edition · Saturday, July 25, 2026Published at 1:43 AM EDT · New York
Researchers Say a Kimi K3 Agent Swarm Found Redis Code-Execution Flaws in 27 Minutes
The 32-agent run on Moonshot AI's 2.8-trillion-parameter model cloned, fuzzed, and built a working exploit against several Redis versions, though the primary flaw requires an authenticated client.
Security researchers report that an autonomous agent swarm built on Moonshot AI's newly released Kimi K3 found remote-code-execution flaws in Redis and assembled a working exploit in as little as 27 minutes. The run, described first in Russian-language coverage and then by English-language outlets, had the agents clone the Redis source code, run fuzzing, and debug crashes in the GNU Debugger inside an authorized test environment.
The technical findings are specific. According to The Hacker News, the work combined a stream consumer-group shared-NACK double-free, tracked as CVE-2026-25589, with a separate heap overflow in the bundled RedisBloom TDigest module, affecting stock builds of Redis 6.2.22, 7.4.9, 8.6.4, and 8.8.0. The double-free lets the same block of memory be freed twice, which gives a reliable method for executing code.
Two caveats limit the alarm. The primary path requires an authenticated client, so it is an escalation after login rather than an unauthenticated break from the open internet. A further claim, that the same system found a zero-click attack chain against Telegram Desktop and iOS, is so far asserted by the researchers rather than independently confirmed.
Kimi K3 is a sparse mixture-of-experts model of about 2.8 trillion parameters that routes each token through 16 of 896 experts, with a one-million-token context and a variant, K3 Swarm Max, built for exactly this kind of large-scale parallel agent work. The significant point for defenders is that a downloadable Chinese model is the foundation for vulnerability discovery at machine speed.
- If true, who benefits
Moonshot AI, whose downloadable Kimi K3 gains credibility as an offense-capable frontier model, and the researchers and cyber-tooling vendors publicizing agent-speed exploitation.
- The nuance
The named CVEs and affected versions recur across outlets, but the 27-minute timing and degree of autonomy are self-reported, the primary path needs an authenticated client, and the claimed Telegram and iOS zero-click chain is unconfirmed.
An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.
What this means
Agentic vulnerability discovery is moving from demonstration to repeatable workflow, and the exposed parties are the maintainers of widely deployed infrastructure such as Redis and the defenders who patch on human timelines. The mechanism is a difference in speed. A swarm that fuzzes, triages, and builds an exploit in under an hour shortens the time between disclosure and exploitation. Because the foundation is an open-weight Chinese model, the same capability is available to any actor who can rent enough computing power, not only well-resourced labs.
What to watch
- Independent confirmation or retraction of the claimed Telegram and iOS zero-click chain, which separates a credible offense-automation result from a marketing claim.
- Whether Redis maintainers issue fixes for the named versions and how fast, a test of whether patch cadence can keep up with agent-speed discovery.
- Adoption of the same swarm approach by defensive teams for pre-release fuzzing, which would show the capability serving defense as well as offense.
Observations to monitor, not financial advice.
Synthesized from: Polylog editors · The Hacker News · Cyber Security News
Part of a tracked trend
Autonomous Agents Move Into Cyber Offense
AI agents increasingly run end-to-end intrusions, chaining supply-chain footholds into privilege escalation and credential theft at machine speed, outpacing human and current automated defenses.
More from this edition
- Anthropic Releases Claude Opus 5, Holding Price Flat While Claiming a Coding-Benchmark Jump
- OpenAI and Apollo Research Publish a Method to Detect Hidden Reward-Seeking in Models
- South Korea Commits to Roughly 260,000 Nvidia GPUs for Sovereign AI
- Anthropic Doubles Its AI-Policy Donation to $40 Million Ahead of US Midterms
- Meta's Brain2Qwerty Decodes Typed Sentences From Non-Invasive Brain Scans at 61 Percent Word Accuracy
- Meta's Open Models Cut a Month of DOE Beamline Analysis to Minutes
- A Utah Copper Mine Adds Boston Dynamics Robots to a Fully Autonomous Operation
- MoE Interpretability Papers Probe How Expert Routing Encodes Knowledge and Frequency
- Meta Launches Muse Media Models Aimed at Editable, Production-Ready Output
- New Study Extracts LLMs' Implicit Theories of What Makes Writing Good
- Musk Says AI Will Soon Outstrip Humans by More Than the Human-Chimpanzee Gap