UK Public Sector Code by Organisation

Inspect: A framework for large language model evaluations

Collection of evals for Inspect AI

ControlArena is a collection of settings, model organisms and protocols - for running control experiments.

Extract residual-stream activations and apply steering vectors (including activation oracles) to any vLLM model during inference.

An Inspect extension for agentic cyber evaluations

A Kubernetes sandbox environment for use with inspect_ai

Reproducing "Natural Emergent Misalignment from Reward Hacking" (MacDiarmid et al., Anthropic 2025) with open-source models. Includes reward-hackable RL environments, misalignment evaluations, training configs, and evaluation scripts. Models trained on OLMo (7B, 32B) and GPT-OSS (20B, 120B).

Optimal stopping package for LLM evaluations

A Proxmox sandbox environment for use with inspect_ai

An EC2 sandbox environment for use with inspect_ai

Report Official Development Assistance (RODA)

This repository created for public search portal

Lie Detectors for the Did you lie? paper

Application code repository for ESOS (Energy Savings Opportunity Scheme)

A proof-of-concept eval testing whether models continue with (or mention) pre-filled malign behaviours.

Granting Authority Schema Service

Source code report for METS (Manage your UK Emissions Trading Scheme reporting service)

Application code for the UK Emissions Trading Scheme Registry

Bulk upload of subsidery awards into data base by using excel file

Climate Change Agreements service application code

Isambard Containers

CHPQA Application Repository

Testing the content, logic and interactivity of the DAIEF self-assessment tool

Analysis code and pseudonymised data for "Conversational AI increases political knowledge as effectively as self-directed internet (arXiv:2509.05219). UK AI Security Institute.

A Python library for running AI debate experiments — orchestrate LLM debaters, judges, datasets, and protocols, and measure whether debate improves AI judge accuracy.