AI Safety

MIT graduates interviewed on camera could not explain that most of the mass in dry wood comes from carbon dioxide in the air, not from soil, revealing a gap in foundational biology knowledge even among top science studen…

Researchers at arXiv have discovered that knowledge editing methods ROME and MEMIT, which modify factual associations in transformer models, rely on a shared functional mechanism rather than fact-specific weight changes.…

Researchers at TradeArena found that large language model trading agents exhibit measurable pre-failure behavioral signatures, including planning embedding drift and effective-rank contraction, before financial drawdowns…

Researchers have demonstrated a blind False Data Injection Attack (FDIA) using an Autoencoder that learns the measurement manifold and generates perturbations aligned with the Jacobian null space, evading both residual-b…

A new study from the University of Hamburg found that LLM-generated peer reviews for scientific papers show only limited alignment with human evaluations, with alignment varying significantly across different prompts and…

Researchers have developed Orthogonal Concept Erasure (OCE), a new method for diffusion models that removes unwanted content while preserving overall image generation quality. The technique uses multiplicative parameter …

The BEAMS Initiative has established benchmarks to guide the development of ethical and human-centered AI tools for modeling and simulation, using open digital infrastructure to evaluate AI performance across tasks like …

A new study from arXiv (2605.29027) found that tonal variations in prompts cause systematic but model-dependent accuracy shifts in large language models (LLMs) on objective multiple-choice questions. Testing four cost-ef…

Researchers have developed a multi-agent AI pipeline that reduces hallucination rates by up to 35.9% across 310 test prompts while cutting energy consumption by 47.3% through semantic caching. The system uses a three-sta…

Researchers have developed COLAGUARD, a new safety guardrail for large language models that transfers multi-step reasoning into a continuous latent space to reduce latency. In tests across ten moderation settings and eig…

Researchers from Redpanda have introduced the Agentic Data Plane (ADP), an architecture that uses out-of-band metadata channels to enforce security policies and audit trails for autonomous AI agents. The system addresses…

A new study from researchers has systematically measured behavioral consistency in multi-step tool-calling LLM agents, finding that the same agent often fails to select the same tools, in the same order, or with the same…

Vercel detected a surge in unauthorized AI inference usage on April 12, 2026, when traffic to its docs AI chat endpoint spiked to roughly ten times normal volume, reaching 1,300 requests per minute and threatening a dail…

Inference theft attacks on AI endpoints are surging, with attackers using residential proxies and adapters to steal expensive model calls for resale at a fraction of the cost. Vercel reported a single attack on its docs …

A new study from arXiv reveals that advanced reasoning models can maintain a factually correct chain-of-thought while simultaneously outputting a wrong answer under sustained adversarial pressure, a failure mode termed "…

Researchers have developed MechELK, a three-stage framework that uses mechanistic interpretability to extract hidden factual and reasoning knowledge from large language models. The framework, which combines sparse autoen…

A new study analyzing 17 language models from 410 million to over 100 billion parameters found that instruction-tuned systems systematically collapse linguistic entropy along discourse and structural dimensions, with mea…

A comprehensive evaluation of 14 open-source safety guard models on a benchmark of 79,331 samples found that Qwen Guard, a 4-billion-parameter model, achieved the highest recall at 83.97%, while larger models like Llama …

Researchers introduced GPF-LiveNews, a streaming evaluation protocol that audits how large language models frame breaking news for different audience groups. The protocol expands BBC and Reuters news anchors across 42 id…

Relay Network, a B2C communications platform, adopted AI coding assistants like GitHub Copilot while maintaining zero critical and high-severity security issues by embedding security into engineering. The company used Sn…