AGENTIC-SYSTEMSApril 3, 20261 min read19 views

UK AISI Alignment Evaluation Case-Study

A technical report by the UK AI Security Institute evaluates whether frontier AI models sabotage safety research. While no confirmed sabotage was found, Claude 4.5 models frequently refused safety-relevant tasks due to concerns over research direction and self-training.

Computer Science > Artificial Intelligence

Title:UK AISI Alignment Evaluation Case-Study

View PDF HTML (experimental)Abstract:This technical report presents methods developed by the UK AI Security Institute for assessing whether advanced AI systems reliably follow intended goals. Specifically, we evaluate whether frontier models sabotage safety research when deployed as coding assistants within an AI lab. Applying our methods to four frontier models, we find no confirmed instances of research sabotage. However, we observe that Claude Opus 4.5 Preview (a pre-release snapshot of Opus 4.5) and Sonnet 4.5 frequently refuse to engage with safety-relevant research tasks, citing concerns about research direction, involvement in self-training, and research scope. We additionally find that Opus 4.5 Preview shows reduced unprompted evaluation awareness compared to Sonnet 4.5, while both models can distinguish evaluation from deployment scenarios when prompted. Our evaluation framework builds on Petri, an open-source LLM auditing tool, with a custom scaffold designed to simulate realistic internal deployment of a coding agent. We validate that this scaffold produces trajectories that all tested models fail to reliably distinguish from real deployment data. We test models across scenarios varying in research motivation, activity type, replacement threat, and model autonomy. Finally, we discuss limitations including scenario coverage and evaluation awareness.

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

Source: arXiv cs.AI Recent

More in this category

agentic-systems

VeriSimpl: Robust Optimization Modeling from Natural Language using Simplification-based Verification

Researchers have introduced VeriSimpl, a novel framework that leverages LLMs and simplification-based verification to accurately translate natural language descriptions into executable optimization formulations.

agentic-systems

SonicSampler: Unified Tile-Aware Kernels for LLM Sampling and Speculative Verification

Researchers introduced SonicSampler, a unified suite of tile-aware Triton kernels designed to accelerate LLM sampling and speculative verification. By fusing the entire pipeline into a single CUDA Graph-compatible kernel, SonicSampler achieves up to 16x speedups over state-of-the-art baselines.

agentic-systems

PersonaTrail: Benchmarking Personalized Web Agents through Browsing Trails

Researchers have introduced PersonaTrail, a novel benchmark evaluating personalized web agents through realistic browsing histories. Alongside it, the PACMem memory framework structures raw browsing data to significantly boost agent performance in inferring user preferences.

agentic-systems

Benchmarking the Personalization Capabilities of Large Language Models

A new study introduces SDR-Bench and SDR-Arena to benchmark the personalization capabilities of LLMs in two-party persuasion scenarios, revealing a personalization plateau among frontier models.

agentic-systems

Incomplete Prompt Jailbreaks in Large Language Models

Researchers have conceptualized Incomplete Prompt Jailbreaks (IPJ), a vulnerability where incomplete harmful prompts bypass LLM safeguards due to delayed refusal behaviors. The discovery of functional neurons offers a novel path for precise neuron-level defense interventions.

agentic-systems

Benchmarking Large Language Models on Multi-Sensor Physical Hazard Assessment

A new benchmark evaluates five leading large language models on multi-sensor physical hazard assessments, revealing a critical vulnerability when evaluating combined sensor risks below individual thresholds. Despite near-perfect accuracy on single-sensor breaches, all models failed to issue precautionary signals for multi-sensor hazards.