Safety · Concept
GuardianBench: Same-Scene Instruction-Contrastive Embodied Safety Benchmark
Instruction-contrastive embodied safety benchmark where benign scenes become hazardous only under unsafe instructions; VLMs average 24.1% pair accuracy.
Forecast horizon
Post-training objectives like VLOS may become standard for instruction-binding safety.
Connections
Connections · 5
How this node ties into the rest of the map, and the evidence behind each link.
to · constrains 4/10
Claude for Physical AI (UST Deployment) Latent instruction-scene hazards benchmark the safety needs of embodied and physical AI agents.
+4 growth to · informs 3/10
Risks of Mass AI Agent Interactions GuardianBench isolates latent instruction-scene compositional hazards relevant to deployed embodied agent safety.
+3 growth to · informs 3/10
Explicit Assessable Understanding in Frontier AI Safety Decisions Latent contextual embodied failures underscore need for explicit safety understanding beyond surface approvals.
+3 growth to · supports 3/10
ASR Error Safety Risks in Voice-Controlled Embodied AI Both isolate latent embodied hazards where instruction or perception context flips a safe scene into harm.
+3 growth from · references 3/10
arXiv AI Safety Feed GuardianBench preprint appears on the arXiv AI safety intake for the period.
+3 growthSignal sources
Signal sources
Dated facts from primary sources in this direction.
US evaluation centre Jun 2025
In June 2025 the US AI Safety Institute was renamed the Center for AI Standards and Innovation (CAISI), pivoting toward security, standards and adversary-model assessment.
NIST → Frontier safeguards May 2025
Anthropic activated its ASL-3 deployment and security standard with Claude Opus 4 on 22 May 2025 — the first real-world trigger of a responsible-scaling tier, focused on blocking bio-weapon uplift.
Anthropic → Cross-border testing 2025
The International Network of AI Safety Institutes (launched Nov 2024) ran a third joint testing exercise focused on agentic AI systems across cyber and fraud strands.
European Commission — AI Office →