Module SDD-02 — RedTeamLLM

RedTeamLLM

Plan correction, memory management, and context-window constraints via the summarize-reason-act loop.

60
minutes
8
artifacts
3
sub-sections
The deep study of coherence on long offensive engagements. RedTeamLLM decomposes the monolithic reasoning call into summarize, reason, act — compressing context before planning, correcting plans on failure, and persisting findings outside the context window. The definitive answer to the three failure modes that break every long-running offensive agent.
Key Claims
Load-Bearing Claims

A monolithic ReAct loop fails on long offensive engagements for three compounding reasons. Context overflow loses information. Plan drift buries the objective in noise. Memory loss discards early findings. RedTeamLLM's summarize-reason-act decomposition addresses all three.

Plan correction is the reason step's job, not the model's implicit responsibility. The reason step sees the prior action's outcome and the summarized state, and revises the plan on failure. This is the explicit handling of error compounding that Course 1 Module 1 flags.

Persistent finding memory must be separate from context. Findings recorded at step 5 must still be available at step 50. Memory inside the context window is memory that summarization can destroy.

The three-step decomposition introduces a novel injection surface. Adversarial content that survives the summarize step reaches the planner. More structure means more surfaces — the tradeoff of decomposition that any decomposed offensive harness must audit.

After This Module
01
Explain why a monolithic ReAct loop fails on long offensive engagements and how summarize-reason-act addresses context overflow, plan drift, and memory loss.
02
Map RedTeamLLM's architecture: the three-step pipeline, persistent finding memory, and plan-correction mechanism.
03
Distinguish RedTeamLLM's depth from CAI's breadth and articulate why the hybrid is the strongest architecture in the roster.
04
Score RedTeamLLM on the 12-module rubric (41/60) and identify the summarize-step injection surface as the novel risk.
05
Implement a minimal summarize-reason-act loop and compare it against a monolithic loop on a 20+ iteration task.
Artifacts