Jul 24, 2026

A new exploit toolkit chains three authenticated remote code execution vectors against recent Redis releases and the bundled RedisBloom module. The attack uses EVAL, RESTORE, and XGROUP commands—already available to authenticated users—to trigger a stream consumer-group double-free, a TDigest heap overflow, and a TopK wild-free. Two of the exploits explicitly bypass prior patches for CVE-2026-25243 and CVE-2026-25589 by exploiting incomplete pointer zeroing in TopK_Destroy.

Anthropic now reports that Claude writes more than 80% of the lines merged into production, up from low single digits in early 2025. The shift is outpacing traditional peer review processes that still assume a human line-by-line reviewer for every change. A parallel analysis argues that while AI may cheapen implementation, the engineering judgment needed to decide what to build, define security boundaries, and evaluate generated code remains as expensive as ever, creating a structural mismatch between generation speed and meaningful review.

Researchers have demonstrated LeakyLMs, side-channel attacks that extract proprietary architecture details from production LLMs using only token-generation timing observed through remote APIs. They can detect speculative decoding and pinpoint the draft model’s context length—for example, confirming that Google Gemini Flash 2.5 uses speculative decoding with a roughly 128K-token draft window. A second technique models token-generation latency to uncover further deployment details.

Other security findings include a Fastly CDN caching flaw on RubyGems.org that could serve one user’s newly issued API key to another person hitting the same edge node within an hour, affecting roughly 18% of sign-ins using gem clients older than v3.2.0. Meanwhile, a teardown of Hanwha Vision camera firmware revealed a hardcoded GitHub admin token baked into dozens of files due to the Vite build process, granting access to hundreds of repositories.

On the infrastructure side, the open-source modern-fs-benchmark evaluates btrfs, ZFS, bcachefs, ext4, and XFS across 26 multi-device configurations, testing silent-corruption recovery, snapshot aging, and rebuild behavior; checksumming filesystems successfully repaired deliberately corrupted copies while classic md/LVM stacks silently returned bad data. Separately, Quantprobe demonstrated that a 110B-parameter GLM-4.5-Air model can stream from a SATA SSD on a 16GB desktop at 0.19 tok/s, while a hybrid VRAM/RAM layout pushed a 30B model to 19.3 tok/s, both matching pre-registered predictions.

Finally, a study on DreamerV3-family agents reveals that catastrophic forgetting in sequential RL tasks stems from a channel problem: the world model retains nearly everything about past tasks, but the actor’s behavior collapses. In a different corner of software engineering, the mp-units library published a type-safe design that separates logarithmic “levels” (e.g., dBm, anchored to a reference) from “gains” (dB, Np), preventing arithmetic errors like adding two dBm values.