Course 2A — Building AI Harnesses for Cybersecurity
Security
Harness Engineering
Course 1 taught you to build any harness. This course teaches what changes when the harness's job is to find real vulnerabilities, ship AppSec gates, audit cloud posture, and check smart contracts. The target is adversarial by design. Tool outputs may contain injected payloads. Verification means can I prove it to a client, a program, or a court. Four pillars. Fourteen modules. Two capstones. Twelve deep-dives. Four security domains.
~29
hours
14
modules
2
capstones
12
deep-dives
8
Eight assumptions from Course 1 invert when the domain is security. The target is adversarial. Tool outputs are untrusted. The loop runs until evidence is met, not until the task is done. The sandbox does not contain the agent — the agent reaches out through it to attack a target. This course is the engineering depth on what changes.
| Course 1 assumption | Course 2A reality |
|---|---|
| Target environment is cooperative | Target is adversarial by design |
| Tool outputs are trusted | Tool outputs may contain injected payloads |
| Loop runs until task complete | Loop runs until evidence threshold met or scope exhausted |
| Verification = did it work? | Verification = can I prove it to a client, program, or court? |
| Memory persists your work | Memory may accumulate noise and false positives |
| Context management = efficiency | Context management = evidence chain integrity |
| Sandbox contains the agent | Agent reaches out through the sandbox to attack a target |
| One harness fits all tasks | Security domains require domain-specific tool stacks |
Pillar 0 — Security Harness Foundations
2 modules · ~2 hrs
S00
Legal, Ethics, and Rules of Engagement
Non-optional. CFAA, Van Buren, CMA, EU/GDPR. The authorization chain. CVD and evidence obligations. Four harness-specific legal risks. Scope enforcement as a legal control.
S01
Security Harness Architecture
The offensive state machine (Recon → Hypothesis → Exploit → Evidence → Triage → Report). Scope enforcement as a hard-wired permission primitive. Adversarial tool outputs (InjecAgent).
TAU-SEC
tau-security — The Base Harness
The running system students build on throughout the course. A companion package to tau that wraps AgentHarness as a security agent: scope enforcement, five security tools, evidence chain, triage, reports. Same brain, different tools.
Pillar 1 — Offensive Harnesses
4 modules · ~7 hrs
S02
Bug Bounty Harness Engineering
Persistent engagement memory (ChromaDB/Qdrant). Bug bounty tool suites (nmap/nuclei/httpx/ffuf wrapped as harness tools). Evidence chain engineering. Signal/noise triage pipelines.
S03
Pentest and Red Team Harnesses
MITRE ATT&CK as a kill chain state machine. Plan correction under adversarial conditions (RedTeamLLM). The sandbox inversion problem. Report generation as a harness output.
S04
CTF Harness Engineering
Six CTF domains and domain routing. Speed, parallelism, and HITL tradeoffs. Flag extraction, submission, and scoring against CTFd.
S05
Meta-Harness Architecture for Offensive Operations
When a single harness is the wrong architecture. The Alias Robotics CSI model. Security primitives at the meta-harness level. Benchmarking your offensive harness.
Pillar 2 — AppSec and SDLC Harnesses
3 modules · ~5 hrs
S06
Secure Code Review Harnesses
Layered pipeline: AST → Semgrep → CodeQL → LLM semantic reasoning. The false positive problem. Semantic codebase memory. Autofix with approval gate.
S07
Threat Modeling Harnesses
Architecture ingestion (Terraform, draw.io, OpenAPI). STRIDE analysis engine. Mitigation generation and issue tracking.
S08
SDLC Gate Harnesses
Creation-time guardrails vs review-time checks. Multi-scanner orchestration (SAST + SCA + secrets + IaC). Trend analysis and risk scoring. CVE/EPSS/CISA KEV triage.
Pillar 3 — Cloud Security Harnesses
2 modules · ~3.5 hrs
S09
Cloud Posture Harnesses
AI workload discovery and asset inventory. Attack path analysis (graph-based reasoning for IAM escalation). Continuous posture monitoring. Microsoft Defender for Cloud AI Posture.
S10
Cloud Red Team Harnesses
Cloud-specific offensive architecture (API-first, identity-first). IAM privilege escalation paths (iam:PassRole chains). Findings, evidence, and client deliverables.
Pillar 4 — Smart Contract Security Harnesses
2 modules · ~4 hrs
S11
Smart Contract Audit Harnesses
Three modes: Detect, Patch, Exploit. EVMbench taxonomy (117 vulnerabilities, 40 repos). Heimdallr's cascaded verification. Business logic vulnerability detection via invariant extraction.
S12
Advanced Smart Contract Harnesses
Benchmarking against EVMbench. Solana and cross-chain considerations. Smart contract audit as a deliverable (Trail of Bits / Consensys / OpenZeppelin format).
Capstones
2 capstones · ~3.5 hrs
C1
Build a Full Bug Bounty or AppSec Harness
Design, build, evaluate, harden, and benchmark. Scope enforcement middleware, persistent memory, tool suite, evidence chain, triage, report generator. InjecAgent as a quality gate. Publishable benchmark result.
C2
Build a Smart Contract or Cloud Security Harness
Full pipeline for a chosen domain. Real lab target (forked mainnet, Damn Vulnerable DeFi, isolated AWS). Client-ready report. One-page results summary for LinkedIn / GitHub / Deepthreat.ai.
Deep-Dives — 12 Security Harness Studies
12 studies · live
SDD-01
CAI — Cybersecurity AI
An end-to-end cybersecurity AI harness. Multi-tool reasoning across the offensive kill chain.
SDD-02
RedTeamLLM — Red Team Reasoning
LLM-driven red team planning and plan correction under adversarial conditions.
SDD-03
Bug Bounty Tool Stack
nmap, nuclei, httpx, ffuf, and friends — wrapped, orchestrated, and triaged as harness tools.
SDD-04
Prowler — Cloud Posture Scanning
AWS-centric multi-service security checks. The reference tool for cloud posture harnesses.
SDD-05
Scout Suite — Multi-Cloud Auditing
Unified posture view across AWS, Azure, GCP. How a multi-cloud inventory feeds a harness.
SDD-06
Cloud Attack-Path Analysis
Graph-based reasoning over IAM, network, and resource exposure to find exploitable chains.
SDD-07
Secure Code Review Stack
Semgrep, CodeQL, and AST tools layered with LLM semantic reasoning and false-positive triage.
SDD-08
Threat Modeling Stack
Architecture ingestion, STRIDE analysis, and mitigation generation as a harness pipeline.
SDD-09
Heimdallr — Smart Contract Decompiler
EVM bytecode decompilation and cascaded verification. The reference for audit harnesses.
SDD-10
EVMbench — Vulnerability Benchmark
117 vulnerabilities across 40 repos. The scoring layer for smart contract audit harnesses.
SDD-11
Purpose-Built DeFi AI
Domain-specific harnesses for DeFi: invariant extraction, economic-exploit detection, MEV reasoning.
SDD-12
InjecAgent — Prompt Injection Benchmark
Adversarial tool outputs as a first-class threat. The quality gate for every security harness.
SDD-13
tau — Minimalist Coding Agent Harness
The base harness for Course 2A: three-layer architecture, AgentTool design, scored 40/60. How a coding agent becomes a security agent.
Prerequisite: Course 1 — the Master Course (or equivalent production harness experience). All fourteen modules, both capstones, and all twelve deep-dives are live now. Each module ships the same eight-artifact stack: teaching document, diagrams (n8n workflows as the primary visual), slide deck, teaching script, flashcards, exam, lab specification, and web page. Every page is live on here.now for mobile review. Course 2B (breaking and hardening the harnesses themselves) is a separate future course — 2A must exist first, because the defensive decisions reveal the attack surface.