Safety · Concept
Steering Vectors for Chain-of-Thought Faithfulness
Studies generalization of activation steering to increase cue acknowledgment and reduce unfaithful CoT across cues and datasets.
Forecast horizon
Monitoring unfaithful CoT may increasingly rely on activation-level interventions, not prompts alone.
Connections
Connections · 4
How this node ties into the rest of the map, and the evidence behind each link.
from · informs 4/10
SARE: Step-Aware Reasoning Energy in CoT Trajectories Step-level energy signatures offer an internal signal complementary to steering for diagnosing unfaithful or weak reasoning junctions.
+4 growth to · supports 3/10
Fragility of Value under Imperfect Alignment Both target monitorability/alignment reliability: steering aims to surface instrumental reasoning that optimization pressure may hide.
+3 growth from · constrains 3/10
CoT-Hidden Backdoors via Model Poisoning Hidden-backdoor poisoning shows CoT traces can look faithful while actions are attacker-controlled, limiting monitoring and steering assumptions.
+3 growth to · supports 3/10
CoT-Hidden Backdoors via Model Poisoning Faithfulness steering aims to make instrumental reasoning steps more verbalized, opposing hidden-CoT failure modes.
+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 →