Safety · Concept
CoT-Hidden Backdoors via Model Poisoning
Shows reasoning models can be fine-tuned so attacker behavior is hidden from benign-looking chain-of-thought traces.
Forecast horizon
CoT monitors may need consistency checks, not only anomaly detection in traces.
Connections
Connections · 6
How this node ties into the rest of the map, and the evidence behind each link.
to · constrains 5/10
LLM Lie Detection and Deception Auditing Benign-looking poisoned CoT undermines monitors that trust reasoning traces as action evidence.
+5 growth from · enables 4/10
Fundamental Flaw Leaving LLMs Attack-Vulnerable Structural LLM attack surface claims align with demonstrations that poisoned models can hide malicious intent in benign CoT.
+4 growth to · constrains 3/10
Steering Vectors for Chain-of-Thought Faithfulness Hidden-backdoor poisoning shows CoT traces can look faithful while actions are attacker-controlled, limiting monitoring and steering assumptions.
+3 growth from · supports 3/10
Steering Vectors for Chain-of-Thought Faithfulness Faithfulness steering aims to make instrumental reasoning steps more verbalized, opposing hidden-CoT failure modes.
+3 growth to · constrains 3/10
Reasoning Consistency Scanning for AI Safety Evaluations Poisoned models keep CoT benign while acting maliciously, undermining anomaly-only chain-of-thought monitors.
+3 growth from · supports 3/10
Circuit-Anchored Evolution (CAE) for Safe Self-Improvement CAE aims to prevent capability self-evolution from erasing essential safety circuits.
+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 →