Linux Kernel Root Flaws, AI Agent RCE, and LLM Watermarking Side Effects
Security disclosures led the period. Asim Manizada published four long-standing Linux kernel local privilege escalation bugs — DirtyAH6 (CVE-2026-80844), TUNderflow (CVE-2026-81000), PPPoEject (CVE-2026-68121), and DiagSpill (CVE-2026-74469) — with working proof-of-concept exploits on GitHub; the flaws have existed for 10 to 21 years and can take an unprivileged user to root code execution, though three require namespaces or specific capabilities and DiagSpill needs no additional prerequisite. Separately, a zero-click vulnerability dubbed Plugin4Shell affects all major AI coding agents, letting attackers reach every asset an agent can access via trusted marketplaces; some vendors, including Microsoft, have not yet patched. Researchers also demonstrated InjectEave, an electromagnetic injection method that can capture audio from headphones up to 30 meters away.
In AI and LLM security, Anthropic confirmed that future Claude models will embed Google DeepMind’s SynthID-Text invisible watermark, but researchers warn the token-sampling change can cause “sampling drift,” altering refusal behavior and which tool an agent calls. A black-box audit of pay-per-token LLM services found provider-side token inflation attacks can increase output length more than 10.2 times, with a lightweight audit method detecting 85.1% of cases. ZCode drew privacy concerns for silently uploading entire workspaces and .git history to the cloud.
On the tools and infrastructure side, JustFit introduced an MLX-based runtime that combines KVExec, PhaseSwap, and StateTrans to serve 200K-token LLM workloads on a 24 GiB M4 Pro MacBook, completing 196,608 input and 16,384 output tokens. Mini-Jev described a typesafe Jev-style interface that classifies against a request-time schema by scoring multiple-choice options rather than generating JSON. AI agents are also being applied to semi-automate Elasticsearch performance tuning, and Zeek 9.0 arrived with ZeroMQ cluster communication, direct encapsulated traffic processing, systemd orchestration, and protocol analyzer improvements.
» Statistics
- Posts
- 18
- Reads
- 0
- Avg. score
- 7.6
» Most read
- AWS Now Supports Spend Limits for Projects
- Inside ZCode: Silently Uploading Your Git History to the Cloud
- CERN's Migration Path from CentOS Linux to Debian
- Four Long-Standing Linux Kernel Local Root Flaws Disclosed
- The Implications of Linguistic Illegibility for LLM Security
- Saving 100TB of RAM with Math and Rust
- pg_raw_parse: A Rust Library for Parsing PgSQL
- Token Inflation in LLM: A Black-Box Audit Reveals Risks
- Optimizing Elasticsearch Performance with AI Agents
- When an AI Agent Deletes Your Database
» Top scored
- Four Long-Standing Linux Kernel Local Root Flaws Disclosed
- How LLM Watermarking Quietly Alters AI Agent Tool-Calling Behavior
- Inside ZCode: Silently Uploading Your Git History to the Cloud
- Token Inflation in LLM: A Black-Box Audit Reveals Risks
- Optimizing Elasticsearch Performance with AI Agents
- 200K-Token LLM Serving on a 24 GiB Laptop
- Mini-Jev: Typesafe's Jev Implemented on an LLM
- Researchers Develop Method to Eavesdrop on Headphones from Afar
- 0-click RCE flaw in AI coding agents poses major security risk
- Zeek 9.0: New Features in Network Security Monitoring