Redis RCE and XFS Local Root Escalation Lead a Week of Security Research; AI Exploit Benchmark and MCP Audit Follow
A new exploit toolkit demonstrated authenticated remote code execution against multiple Redis versions (6.2.22, 7.4.9, 8.6.4, 8.8.0, 8.8.1), chaining a double-free in stream consumer groups, a TDigest heap overflow, and a wild-free bug in the bundled RedisBloom module. Two of the exploits are direct patch bypasses of CVE-2026-25243 and CVE-2026-25589, requiring only authenticated commands like EVAL, RESTORE, and XGROUP—making upgrades urgent for Redis deployments.
Qualys disclosed RefluXFS (CVE-2026-64600), a race condition in the Linux kernel's XFS filesystem that lets an unprivileged local user overwrite any readable file, including root-owned config files and SUID binaries, on any XFS volume created with reflink support—the default since 2019. The attack reflink-clones a target into a scratch file and races concurrent O_DIRECT writes, exploiting a lock‑drop window in the copy‑on‑write allocation path.
Researchers from UC Berkeley, MPI, Anthropic, OpenAI, and Google released ExploitGym, a benchmark of 869 real‑world vulnerabilities across userspace, Chrome's V8, and the Linux kernel. AI agents are given a crash‑triggering input and must achieve unauthorized code execution; standard mitigations like ASLR, stack canaries and V8's heap sandbox reduce but don't eliminate success, with agents finding partial‑pointer overwrites and sandbox escapes.
The AI security ecosystem also faced scrutiny. A static analysis tool called mcpgrade measured how well AI agents can use MCP server tools by checking parameter descriptions; 11 of 36 popular servers—including official ones from MongoDB, Notion, Airtable, Todoist, and firecrawl—scored D or F due to missing descriptions from auto‑generated schemas. In a separate development, Anthropic reported that Claude now writes more than 80% of its production code, prompting a redesign of peer review processes that still assume a human reviews every line. Meanwhile, LeakyLMs timing attacks extracted proprietary LLM architecture details remotely, confirming that Google Gemini Flash 2.5 uses speculative decoding with a ∼128K‑token draft window.
A RubyGems CDN caching bug on Fastly can leak legacy API keys across users, affecting about 18% of gem client sign‑ins, including the gem tool bundled with macOS Tahoe. A reverse‑engineering effort documented Windows' undocumented offline device ID API via clipc.dll and clipsvc.dll. Martin Kleppmann's 2014 talk on turning the database inside‑out with Apache Samza and Kafka resurfaced, and a demonstration showed factual knowledge hand‑wired directly into Llama‑3.1‑8B weights using tiny neuron circuits—no fine‑tuning required.
» Statistics
- Posts
- 379
- Reads
- 11
- Avg. score
- 7.8
» Most read
- How We Self-Host DeepSeek V4 Flash on AWS Spot Instances
- Quantprobe Runs a 110B-Parameter LLM on a 16GB RAM 2016 Desktop
- New Linux benchmark tests corruption, snapshots, rebuilds on 26 layouts
- No Fine-Tuning: Facts Hand-Wired Directly Into Llama-3.1-8B's Weights
- GPT-5.6 vs Claude Fable 5: Benchmarking Physical AI Simulation
- Kimi Linear: Hybrid Attention Architecture Beats Full Attention
- Reverse-Engineering Windows' Undocumented Offline Device ID API
- Three Search Models in 2026: Keyword, RAG, and Conversational LLM
- StemDeck: Free Local Stem Separation and DAW Mixer
- Inside the New Agent Swarm: Coordination at 1,000 Commits/Sec
» Top scored
- Authenticated RCE Exploits Hit Redis via Stream, TDigest, TopK Bugs
- RefluXFS: XFS reflink race lets Linux users escalate to root (CVE-2026-64600)
- ExploitGym Benchmark Tests If AI Agents Can Build Real Exploits
- mcpgrade audit: a third of 36 popular MCP servers score D or F
- When AI Writes Most of the Code, Peer Review Must Be Redesigned
- LeakyLMs: Timing Attacks Expose LLM Architecture and Inference Tricks
- Reverse-Engineering Windows' Undocumented Offline Device ID API
- Turning the Database Inside Out: Kleppmann's Samza Talk Revisited
- RubyGems CDN Caching Bug Could Leak Legacy API Keys to Other Users
- No Fine-Tuning: Facts Hand-Wired Directly Into Llama-3.1-8B's Weights