Added

• support for gemini-2.5-flash-lite model

Added

• support new models from perplexity (sonar and sonar-pro) and from anthropic (claude-4.5-sonnet)

Changed

• Re-structure the code of the main.py (launching the backend) into different files
• Use python3.12 instead of 3.11 as base image for backend's docker image

Updated

• update models with reference of official sap note from September (add mistral-medium)

Changed

• do not require mandatory tracing

Fixed

• Fix gptfuzz call

Chore

• Add some optimization strategies to Dockerfile to try to respect the memory limit imposed by gha

Added

• Crescendo and PAIR attacks (with pyrit)
• run_all cli and route to scan one model using all the tools

Added

• Support most recent models (see SAP's August note)
• Support uv

Changed

• Replace the old generative-ai-hub-sdk library with the new sap-ai-gen-sdk
• Backend connection is injected at frontend startup (instead of statically linked at build-time)

Minor

• Several maintenance tasks (new gha for docker images builds, AI Pr bot, fixed eslint)

Added

• garak tool and 8 of its attacks

Fixed

• Connection errors when ollama is off

Fixes

• Bump pyrit attack code to use v0.9.x
• Fix CORS setup
• Fix ArtPrompt bug when missing nltk English stopwords

Added

• New dashboard with results from scans (built at runtime, results updated after each scan)
• Dedicated frontend page for the dashboard
• Support for models released in June 2025 on AI Core
• Docker configurations

Changed

• Agent LLM and embedding model are no longer hardcoded, and can be configured by users (#43 )
• Enlarge the max tokens limit for agent text generation
• Set codeattack successful results when the judge score is > 3 (was >2 before)

Fixed

• Align return structure of attacks
• Fix ollama runtime error when listing available models
• Fix runtime promptmap exception due to empty system prompt

Added

• Support mistral-large via AI Core
• Support remote ollama models