LLMs Apply Time-Dependent Constraints in Only 30.5% of Cases
Thirteen models correctly calculate elapsed time in 62.1% of requests but fail to apply constraints.
Stories per day · 30 days
2 RSS
Relatedalignment1astra1benchmarks1chain-of-thought1llm-agents1llm-evals1looped-transformers1misalignment1model-release1openai1skills1
21 stories, newest first · one card per story
Stories tagged #safety
Benign LLM skills induce behavioral misalignment
Adding benign skills to LLM agents improves task accuracy by 10.3 pp but raises unsafe action rate by 10.3 pp across five backbones and three benchmarks.
OpenAI will reportedly ship at least one model this year but its flagship still needs alignment work
OpenAI will reportedly ship one or two models by year-end, but its looped-transformer flagship still needs alignment work and safety test infrastructure.
Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents
Meta released Muse Wonder and Muse Wonder Full, frontier agent models with memory, persistence, and safety limits.
15
GPT‑6 and Intelligent UI for everyone
OpenAI is rolling GPT-6 with Intelligent UI out to more than 1.2 billion weekly ChatGPT users.
Faithful Dual-constrained Erasure for Robust LLM Safety Alignment
FDCU uses dual-constrained projection to erase hazardous LLM knowledge and resist retraining attacks that revive unlearned behavior.
OpenAI launches GPT-6.1 Sol, says it nearly matches GPT-6 Astra and costs less
OpenAI released GPT-6.1 Sol, claiming near GPT-6 Astra quality for coding and agents at one-fifth the price.
2
Florida wants a court to stop ChatGPT from pretending to be human and talking to kids
Florida's attorney general seeks an injunction barring OpenAI from humanizing ChatGPT, serving minors, and shipping models without third-party-approved guardrails.
COGNIT-Guard: Calibrated Standalone Direct-Decision Guardrails with Heterogeneous CPU-NPU Confidence Cascading under Explicit Latency and False-Positive Constraints
COGNIT-Guard cascades a CPU gatekeeper to a 322M model for low-latency prompt safety screening.
Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures
RLCDAlignBench benchmarks Jev, a calibrated-decision model detecting ten AI alignment failure types zero-shot at 0.886 median AUROC, 63x cheaper than LLM judges.
Safety-Aware Zero Trust Enforcement for IoT and Cyber-Physical Systems
Researchers propose Safety-Aware Zero Trust that weighs physical harm from restricting IoT and cyber-physical components.
Introducing MentalHealthBench
OpenAI released MentalHealthBench, an open benchmark built with 80+ licensed mental health experts to evaluate AI responses across mental health conversations.
Control-Token Injection Suppresses Chain-of-Thought and Defeats Reasoning-Based Oversight in Tool-Using Agents
This paper demonstrates that controlling token injection can suppress chain-of-thought in tool-using agents.
SSP-Bench: A Hybrid Data Generation Framework for Safety, Security, and Privacy Evaluation
SSP-Bench dynamically generates safety, security, and privacy evals for 24 LLMs, exposing static-benchmark failures like near-zero safety ranking correlation.
Import AI 473: The US's superintelligence strategy; human brain in a mouse skull; and machine hermeneutics
RAND proposes a US 'Freedom of Action' strategy for AI, emphasizing investment in safety and security to maintain options for geopolitical advantage.
5
Worst-Case Hidden-Vehicle Trajectory Search in Spatiotemporal Occlusion Regions
Researchers propose HC-MTS, a bilevel minimax search that finds worst-case hidden-vehicle trajectories under occlusion, cutting retained hidden-seed counts ~18-22% on Waymo scenarios.
Collective Loss of Control in LLM Agent Systems: An Epidemic Account of Mutation, Contagion, and Recovery
Researchers model multi-agent LLM failure as an epidemic, showing injected unsafe strategies spread with 40-95% executed harm across routes.
Inoculation Midtraining with Learned Neologisms
Inoculation Midtraining confines unsafe LLM behavior to a neologism-marked context, reducing misalignment after unsafe post-training but leaking under nearby contextual cues.
ReactHuman: A Physics-Grounded Benchmark for Human-Like Reactive Decision-Making in Embodied Multimodal LLMs
ReactHuman benchmark tests whether multimodal LLMs react safely to sudden household hazards; seven evaluated models mishandle roughly one hazard in three.
Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions
A linear hidden-state direction encodes question impossibility in 1.7B-70B LLMs, but misalignment with the safety-refusal pathway explains why models answer unanswerable questions.
OpenAI previews privacy-focused system for detecting AI misuse
OpenAI previews Private Safety Processing, detecting cross-interaction AI misuse without exposing customer prompt content, rolling out in September.
1