Sam's News — tech-research — 2026-08-20¶
Neuroscience¶
7.5 Non-Invasive Brain-Computer Interface Decodes Natural Sentences From Brain Recordings¶
Researchers developed Brain2Qwerty v2, a non-invasive brain-computer interface that decodes natural sentences from magnetoencephalography recordings. Trained on 22,000 sentences from nine subjects over 10 hours each, the model achieves 39% average word error rate, with best-performing participants accurately decoding half of sentences with one or fewer word errors. The approach combines deep learning, fine-tuned language models, and AI agents.
- Average word error rate: 39% across participants
- Best participant: 50% of sentences decoded with ≤1 word error
- Training data: 22,000 sentences from 9 subjects, 10 hours per subject
- Uses MEG (magnetoencephalography) non-invasive brain recordings
- Performance improves log-linearly with increased data volume
- Targets communication restoration after brain injury
Sources: arXiv AI Web Searched, arXiv — Computation and Language RSS
AI¶
6.5 Abliteration Mitigation via Refusal Aliases¶
Researchers propose a method to mitigate abliteration attacks that bypass safety alignment in large language models through refusal alias techniques.
Sources: arXiv — Computation and Language RSS, arXiv — Cryptography and Security RSS Update to: Research explores defensive deception against model safety-removal attacks
6.5 Redakto: privacy-preserving layer for LLM interactions¶
Researchers introduce Redakto, a method to redact personally identifiable information from LLM inputs and outputs to preserve user privacy.
Sources: arXiv — Artificial Intelligence RSS, arXiv — Cryptography and Security RSS
Security¶
6.5 CTIFoundry: agent-native framework for cyber threat intelligence¶
A new corpus scaffold enables LLM agents to autonomously conduct multi-step cyber threat investigations at query time.
Sources: arXiv — Artificial Intelligence RSS, arXiv — Cryptography and Security RSS
NLP Research¶
6.5 Backdoor Learning in Language and Vision-Language Models¶
Research examines vulnerabilities to backdoor attacks in large language models and vision-language models.
Sources: arXiv — Computation and Language RSS
AI Safety¶
6.5 LLM Behavior in Medical Allocation Varies With Inference Setup, Not Just Inputs¶
Researchers find that how large language models are prompted and executed—not just what they're asked—significantly affects their medical decision-making behavior.
Sources: arXiv — Computation and Language RSS
Healthcare AI¶
6.5 Fine-Tuned LLM Framework Provides Automated Clinical Supervision and Risk Triage in Mental Health¶
A new tri-stream fine-tuned LLM framework addresses mental healthcare's supervision gap by automating clinical oversight for novice therapists.
Sources: arXiv — Computation and Language RSS
AI Risk¶
6.5 Agentic AI Implementations Expose Enterprises to Integration and Governance Risks¶
According to Info-Tech Research Group, agentic AI systems have evolved from chatbot pilots into autonomous systems that trigger workflows and access sensitive data, creating governance challenges.
Sources: PR Newswire Web Search Update to: Agentic AI deployments pose enterprise integration and governance risks
Geopolitics¶
6.5 Fudan University and Peers Targeted as Universities Become Casualties of US-China Tech War¶
Fudan University and Shanghai Jiao Tong University have been placed on a U.S. national security list while China targets European counterparts in escalating tech tensions.
- Fudan and Shanghai Jiao Tong added to DoD Section 1286 list on July 23, 2026
- DoD-funded researchers prohibited from relevant cooperation with listed institutions
- China placed Wrocław University (Poland) on export control list on July 24 in response
- Measures intended to prevent government-funded research used for military/strategic purposes
Sources: ThinkChina Web Search Update to: Fudan University and Others Targeted in US-China Tech War Escalation
Tech Research¶
6.5 Virginia Tech receives $20 million grant for cloud-based semiconductor lab¶
Virginia Tech has been awarded a $20 million grant to develop a cloud-based semiconductor research laboratory.
Sources: Cardinal News Web Search