Five independent AI vulnerability-discovery-and-remediation harnesses, reviewed side by side — four companion essays, two architecture diagrams, and one consolidated reference list.
Four reviews of what five teams actually built, plus a standalone explainer on what an ideal harness would look like. Start with whichever framing matches how you'll use this.
The full review: what Glasswing is, who built what, component-by-component and architecture-vs-principles analysis, what a finding costs, the metrics stack to track, whether a new harness should be built, and the open-source remediation problem none of the five have solved.
Read the essay →First-person, implementation-focused: what to steal, what to argue about in code review, and where the framework debate (LangChain/LangGraph vs. direct CLI vs. OpenAI Agents SDK) is genuinely unsettled.
Read the essay →Risk, governance, and bug-bounty economics: trust boundaries worth signing off on, the metrics gap you shouldn't cite blind, and how this changes your disclosure and embargo obligations.
Read the essay →Plain language, no engineering background assumed: what these systems are, whether they work, what they cost in practice, the unsolved open-source gap, and five questions to ask your CISO or CTO.
Read the briefing →An ideal harness walked through end to end, with seventeen inline diagrams: context and level-1 DFDs, agent flow, context handoff, agent and tool authentication, block and detailed architecture, trust boundaries, signal fusion, telemetry and SIEM split, sandbox containment, global multi-region topology, four reference stacks (Azure, AWS, GCP, self-hosted), and flow diagrams for code, running systems and configuration. Covers the six target types, agents and loops, memory poisoning, tool access, guardrails, exploit authorization, false-alarm control, remediation per territory, and the export-control question most teams never ask.
Read the explainer →The harness modelled explicitly as a graph: agents and deterministic tools as vertices, handoffs as typed edges that carry trust tier, certainty and provenance but never authority, durable state outside the agents, and policy evaluated on the edge rather than inside the agent that wants to cross it. Includes the handoff envelope schema and its certainty-transition rules, the four human gates, how state is tiered by durability, how short-lived credentials are minted per job, the split between deterministic tools, MCP servers and skills, and Level 1 and 2 data-flow diagrams for code, running systems and configuration.
Read the design →| Harness | Built by | Status | Core mechanism |
|---|---|---|---|
| VVAH Visa Vulnerability Agentic Harness | Visa | Open source (Apache 2.0) | 11-stage pipeline (S0–S11), config-selectable backend per role |
| VulnHunter | Capital One | Open source | 3 Claude Code skills: Hunt → Fix → Verify, zero agent framework |
| STRIX | Independent OSS | Open source | Root/child agent tree, shared Docker sandbox, OpenAI Agents SDK |
| DCRH Defending Code Reference Harness | Anthropic | Open source (reference) | Find → fresh-container grade → patch/re-attack, direct CLI subprocess |
| VDH / VVS Vulnerability Discovery Harness / Vulnerability Validation System | Cloudflare | Internal | 8-stage discovery loop feeding a shared, multi-harness triage system |
| OCR Open Code Review (control, non-Glasswing) | Alibaba | Open source | Code review, not vuln discovery — same deterministic-scaffold pattern, independently arrived at |
The underlying design documentation reviewed for each harness — HLD/LLD, module references, data flows, and operating runbooks. Each hub links every document in its set.
11-stage pipeline (S0–S11) — design, modules, process runbooks, and graph-intelligence tooling.
Read the overview → All docs →Hunt → Fix → Verify skills — design (incl. deterministic-vs-LLM), modules, and the Visa comparison.
Read the overview → All docs →Root/child agent tree — agentic framework design, module reference, and operating processes.
Read the overview → All docs →Reference harness — architecture overview, HLD/LLD, and patch-and-verify runbooks.
Read the overview → All docs →The non-Glasswing control — code-review harness sharing the same deterministic-scaffold pattern.
Read the overview → All docs →