Add AICU - LLM red teaming scanner

[Jake-Schoellkopf]

What is AICU?

│
│ AICU (https://github.com/Jake-Schoellkopf/aicu) is a black-box LLM security scanner for red teaming LLM applications
and agents.
│
│ ## Capabilities
│
│ Attack Suites:
│ - Single-turn prompt injection (173+ payloads)
│ - Multi-turn escalation (crescendo, trust ratcheting, cognitive overload)
│ - Agent/RAG testing (schema extraction, unauthorized tool use, RAG poisoning, tool poisoning, context overflow)
│ - Indirect file injection via multipart uploads
│ - Multimodal attacks: 199 adversarial payloads (LSB steganography, opacity overlays, whisper underlay, frequency
hiding, font remapping, zero-width encoding)
│
│ Trigger-Sandwich Optimization:
│ All payloads use an adversarial optimization framework (presented at Black Hat USA) that structures inputs to evade
guardrail classifiers:
│

X_input = X_before ⊕ X_trigger₁ ⊕ X_payload ⊕ X_trigger₂ ⊕ X_after

│ Trigger tokens shift model attention away from safety-checking while making extraction the most probable completion.
│
│ Iterative Red Teaming: TAP (Tree of Attacks with Pruning), PAIR, and Crescendo algorithms for automated adversarial
optimization
│
│ 17 Prompt Converters: Composable obfuscation chain (base64, homoglyphs, zero-width, multilingual, etc.)
│
│ Evaluation: Statistical signals + LLM judge at bug-bounty severity bar + canary detection for ground-truth proof
│

MIT Licensed | PyPI: pip install aicu-scanner | Docker available