Sam's News — tech-research — 2026-08-27¶
Security¶
7 MIT researchers discover processor security gap exploitable for privilege escalation¶
MIT CSAIL researchers disclosed TONTOU, a processor security vulnerability exploiting a timing gap between when CPU prediction machinery is wiped and when protections take effect. The attack uses interrupt injection to re-contaminate the processor's prediction machinery before defenses activate, successfully tested against Intel and AMD generations and demonstrated extracting a Linux system's password file.
- TONTOU (Time-Of-Check-Time-Of-Use for On-chip Prediction Units)
- Affects Intel and AMD processor generations
- Defeats both software and silicon-based security mechanisms
- Interrupt injection forces execution during vulnerability window
- Successfully extracted Linux system password file
- Undermines Spectre-era CPU security assumptions since 2018
- Researchers: Daniël Trujillo and Mengjia Yan (MIT CSAIL)
Sources: MIT News AI Web Searched, MIT — Computers RSS
AI¶
6 Google releases Gemini Omni 1.1 Flash with expanded control features¶
Google announces Gemini Omni 1.1 Flash, enabling developers to build with more granular control.
Sources: DeepMind Blog RSS
AI Research¶
6 World's first double-blind AI evaluation framework piloted¶
Researchers are piloting the world's first double-blind evaluation methodology for AI systems.
Sources: DeepMind Blog RSS
5.5 Study Traces Persistent Dialectal Biases Throughout the Language Modeling Pipeline¶
Research reveals systematic performance gaps for non-standard dialects persist across all stages of language model development, from tokenization through inference.
Sources: arXiv — Computation and Language RSS
5.5 Multi-Turn Diagnostic Dataset Enables Clinically Interactive LLM Evaluation¶
MTDiag introduces an interactive diagnostic dataset where LLMs incrementally gather information through multi-turn dialogue, more closely simulating clinical practice.
Sources: arXiv — Computation and Language RSS
5.5 ReliableRAG Combats Misinformation in Question Answering via Reliability-Guided Reasoning¶
A framework improves retrieval-augmented generation reliability by guiding reasoning chains to assess and avoid false or misleading retrieved information.
Sources: arXiv — Computation and Language RSS
AI Safety¶
5.5 Study reveals refusal geometry patterns strengthen AI safety against jailbreaks¶
Research finds that diverse refusal prefixes can increase the stability of refusal mechanisms in aligned language models, making them more robust to vector ablation attacks.
Sources: arXiv — Machine Learning RSS, arXiv — Cryptography and Security RSS
5.5 Adaptive Triggering Corrects Demographic Bias in LLM Chain-of-Thought Reasoning¶
A method mitigates stereotype amplification during intermediate LLM reasoning steps where chain-of-thought prompting can expose and reinforce demographic biases.
Sources: arXiv — Computation and Language RSS
Security Research¶
5.5 Backdoor vulnerability found in Vision Mixture-of-Experts architectures¶
Researchers identify a vulnerability in Vision MoE token dispatch mechanisms that enables backdoor attacks through capacity overflow in expert-switching systems.
Sources: arXiv — Computer Vision RSS, arXiv — Cryptography and Security RSS
AI Security¶
5.5 Bit-Flip Attacks on Mixture-of-Experts LLMs Can Trigger Infinite Generation Loops¶
Researchers demonstrate a vulnerability in MoE language models where bit flips in routing mechanisms can be exploited to seed infinite generation loops.
Sources: arXiv — Computation and Language RSS