agent blueprint

Architecture insights from verified research

8 findings that should shape the agent's design, each traced back to a paper or repository in this archive.

Composition Beats Model Selection

Finding: Sailor pipeline found 379 vulns vs Claude Code alone finding 12 (30x)
Why it matters: The agent's architecture (how components compose) matters more than which LLM it uses
How to apply: Design multi-phase pipelines (static analysis → LLM synthesis → symbolic execution → validation), not single-pass LLM calls

Every Component Is Essential

Finding: Removing ANY guidance mechanism collapses performance to near-baseline
Why it matters: Don't build a "good enough" agent with missing pieces — each component (targeting, feedback, validation) is load-bearing
How to apply: Include ALL of: static analysis targeting, iterative LLM feedback loops, concrete validation, evidence capture

Multi-Agent, Multi-Model

Finding: Different LLMs for planning (GPT-o3-mini) vs execution (LLaMA 3.1) vs summarization (GPT-4o) increases accuracy
Why it matters: One model for all tasks is suboptimal
How to apply: Route tasks to specialized agents/models based on task type

Scope Enforcement at OS Level

Finding: Deterministic hooks outside the agent loop prevent scope violations
Why it matters: Autonomous agents need hard boundaries, not prompt-instruction boundaries
How to apply: Implement `scope_check.py`-style hooks that run OUTSIDE the agent loop, CIDR-aware

Evidence-First

Finding: No theoretical findings — only verified PoCs with captured evidence
Why it matters: Bug bounty platforms reject reports without working proofs
How to apply: Every finding must produce: working PoC + HTTP evidence + screenshot + CVSS

Disclosed-Report Patterns

Finding: 681 disclosed-report patterns + 2887 HackerOne cases available as structured data
Why it matters: Agent should learn from what actually gets paid, not just theory
How to apply: Train/fine-tune on disclosed reports; use as few-shot examples for vulnerability classes

MITRE ATT&CK Mapping

Finding: Every technique mapped to ATT&CK, CSF 2.0, ATLAS, D3FEND, AI RMF
Why it matters: Structured threat intelligence enables systematic coverage
How to apply: Map every agent action to ATT&CK technique for coverage tracking

Lifecycle Security

Finding: Protection across initialization → input → memory → decision → execution
Why it matters: Agent security failures propagate across stages
How to apply: Defense-in-depth at each agent lifecycle stage

Phase roadmap

complete

1. Research & curation

This archive: sources, papers, repositories and taxonomies.

2. Architecture design

Agent roles, model routing, tool interface, scope enforcement and evidence capture.

planned

3. Agent implementation

Recon, analysis, validation and reporting stages with deterministic guardrails.

planned

4. Testing & iteration

Benchmarks against known targets, ablation studies and regression suites.

Scope enforcement is not a prompt. The research is consistent: deterministic hooks outside the agent loop define what a target is, what may be touched, and when to stop.

Evidence-first. A finding without a reproducible proof-of-concept and captured evidence is a hypothesis, not a result. The agent must validate before reporting.

Next concrete steps

The archive is Phase 1. The research notes suggest this order for Phase 2:

Read the anchor papers (Sailor, ARACNE, ELFuzz, NOVA, Aegis).
Study the Claude Code agent repositories for orchestration patterns.
Design a multi-agent pipeline with explicit scope and evidence gates.
Convert disclosed-report patterns into structured evaluation data.
Implement recon → analysis → validation → reporting with deterministic limits.

Where to look

Map each design question to the archive:

papers repositories LLM security AI fuzzing methodologies taxonomy