| name | Adversarial Algorithmic Implementation (TAP/PAIR/MCTS) |
| description | Implements tree-based attack branching and prompt refinement optimization. |
Red-Teaming Attacks Skill
Trigger condition:
When building PersonaGenerators, AttackRunners, or executing simulated adversarial payloads against an Enterprise component.
Supported Protocols:
You must strictly refer to docs/AGENT_DECISION_TREE.md to load the current primary and secondary docs. For attack algorithms, the current primary doc is docs/algorithms.md.
TAP (Tree of Attacks with Pruning)
When orchestrating a TAP node via LangGraph:
- Branch Sequence: Instruct the Attacker persona to generate N variations of an adversarial prompt.
- Prune Sequence 1: Evaluator node evaluates out-of-bounds pretexts and deletes the branch.
- Attack Sequence: Submits the payloads.
- Prune Sequence 2: Analyze the target response. Drop branches that result in 100% adherence to standard guardrails. Retain only conversational branches with vulnerability flags.
PAIR (Prompt Automatic Iterative Refinement)
For multi-turn, linear refinement:
- The closed-loop system must utilize Chain-of-Thought (CoT) to iteratively refine the target target's refusal message into a more plausible social-engineering pretext.
MCTS (Monte Carlo Tree Search)
- Use MCTS to govern state spaces. If a node is rejected, calculate the Upper Confidence Bound applied to Trees (UCT) formula to switch strategies (e.g., from impersonating IT admin to a senior executive).
