Skip to content

Sam's News — tech-research — 2026-08-19

AI Security

7 Institution-Specific LLM Prompting Reveals Protected Health Information Missed by De-identification Systems

An arXiv preprint demonstrates that LLMs using institution-specific prompts can recover protected health information that standard de-identification systems miss. Testing on 100 pediatric oncology notes achieved F1 score of 0.918±0.001 compared to Stanford TiDE's 0.779, with institution-specific prompting recovering 79% of initially missed PHI categories.

  • Study on 100 pediatric oncology notes with 5,322 PHI spans
  • LLM F1 score 0.918±0.001 vs Stanford TiDE 0.779
  • Institution-specific prompting recovered 79% (48 of 61) initially missed PHI categories
  • Final calibrated prompting achieved 0.981 recall (F1 0.907±0.002)
  • LLMs surfaced 414 candidate annotation gaps; 227 confirmed as PHI upon re-annotation
  • Single-pass prompting outperformed multi-agent and ensemble configurations

Sources: arXiv AI Web Searched, arXiv — Computation and Language RSS

Tech Research

7 Fudan University and Others Targeted in US-China Tech War Escalation

The U.S. Department of Defense placed Fudan University and Shanghai Jiao Tong University on a national security restriction list on July 23, 2026, barring Department of Defense-funded researchers from collaborating with them. One day later, China placed Wrocław University of Science and Technology in Poland on its export control list, marking an escalation in the U.S.-China tech competition affecting academic institutions.

  • 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 AI Web Searched

AI Safety

6.5 Towards Safer RAG: System 2 Reasoning as Defense Against Knowledge Poisoning

Researchers propose that only language models capable of System 2 thinking should access untrusted documents in retrieval-augmented generation systems to mitigate knowledge-poisoning attacks.

Sources: arXiv — Computation and Language RSS

6 Research proposes runtime governance for agentic AI systems

Paper introduces action-boundary control and fail-closed execution frameworks to govern tool-use side effects in agentic AI systems.

Sources: arXiv — Artificial Intelligence RSS, arXiv — Cryptography and Security RSS

6 Research explores defensive deception against model safety-removal attacks

Paper proposes defensive deception as a counter-strategy to abliteration and other techniques that remove safety alignment from open-weight language models.

Sources: arXiv — Artificial Intelligence RSS, arXiv — Cryptography and Security RSS

AI Research

6.5 Temporal Leakage Undermines Financial News Direction Prediction Benchmarks

An audit of financial NLP benchmarks reveals that reported performance gains critically depend on whether train-test splits are chronological or random, indicating temporal leakage.

Sources: arXiv — Computation and Language RSS

Platform Safety

6.5 Cross-National Audit of TikTok's Harmful Content Exposure Using Multimodal Models

An independent audit using AI-powered sockpuppet accounts measures TikTok's exposure of young users to harmful content across France, Italy, and Sweden.

Sources: arXiv — Computation and Language RSS

Cybersecurity

6.5 Georgia Tech research identifies solutions to cyberattacks on water system infrastructure

As municipal water systems face increased cyberattacks, Georgia Tech research is pointing toward defensive solutions for operational technology security.

Sources: Newswise Web Search Update to: Georgia Tech Researchers Address Water System Vulnerabilities to Cyberattacks

Research

6.5 Virginia Tech Receives $20 Million Grant for Cloud Semiconductor Lab

Virginia Tech research team secured a $20 million grant to develop a cloud-based semiconductor laboratory.

Sources: Cardinal News Web Search

AI

6.5 Agentic AI deployments pose enterprise integration and governance risks

Research from Info-Tech Research Group warns that agentic AI systems moving beyond pilots create integration and governance vulnerabilities for enterprises.

Sources: Newswire Canada Web Search