Proven Exploitable
The Practitioner's Guide to AI-Augmented Vulnerability Discovery
Author: Bindiya Priyadarshini Date: June 2026 DOI: 10.5281/zenodo.21159028 Permanent URL: https://zenodo.org/records/21159028 Licence: Creative Commons Attribution 4.0 International (CC BY 4.0)
The finding
Two AI systems have crossed a capability threshold that changes the economics of vulnerability discovery.
| System | CyberGym Score | Architecture | Status |
|---|---|---|---|
| Microsoft MDASH | 96.5% (any crash) | 100+ agent ensemble | Limited private preview |
| Anthropic Mythos Preview | 83.1% | Single frontier model | — |
| Project Glasswing GPT-5.5-Cyber | 85.6% | Single model | Available |
| Commercial SAST tools | 30–50% | Rule-based | Generally available |
MDASH found 16 confirmed vulnerabilities in Microsoft's May 2026 Patch Tuesday release before they shipped. Four were rated Critical. Commercial SAST tools score 30–50% on the same benchmark these systems are hitting at 83–96%.
The gap between AI-augmented discovery and traditional tooling is not marginal. It is structural.
What this paper covers
The systems
MDASH architecture — five-stage pipeline, 100+ specialised agents, tiered model routing, prepare → scan → validate → dedup → prove Mythos Preview autonomous discovery workflow — attack surface ranking, hypothesis formation, confirmation, PoC generation CyberGym benchmark analysis — what 96.55% actually means and what it doesn't cover
Real-world evidence
May 2026 Patch Tuesday cohort — 16 CVEs found by MDASH before release Four Critical findings documented with CVE IDs, CVSS scores, and vulnerability class TeamPCP supply chain incident analysis
Enterprise integration
Nine-layer reference architecture including new deployment mapping and reachability layer Deployment mapping — which code is deployed where, internet-exposed vs. internal, reachability-adjusted priority Microsoft Sentinel and MCP server integration GitHub Code Security and Copilot Autofix integration Defender advanced hunting integration
Supply chain and coverage
Third-party OSS library scanning Container image and AI model supply chain Language and stack coverage — honest assessment per language
StackEvidence qualityKey risksC / C++Strong — all published resultsKernel, firmware, cryptoJava / JVMQualifiedDeserialisation, auth bypassPythonLimitedAI supply chain, pickle deserGo / RustLimitedLogic bugs onlyWeb stacksUnvalidatedDifferent problem classCloud-nativeSpecial handlingIAM-centric, workload baselinesOT / ICSHigh riskDetection only, no autonomous action
Governance
Scan scope policy Finding ownership and accountability Model update protocol Regulatory controls
8 prioritised recommendations for security engineers and CISOs
Files in this repository
FileDescriptionProvenExploitable_FullPaper.pdfFull paper — real-world CVE evidence, MDASH and Mythos architecture, nine-layer enterprise integration, 8 recommendationsProvenExploitable_QuickReference.pdfTwo-page standalone quick reference — benchmark table, 8 actions, architecture layers, language coverage, governance checklistAI_Security_Architecture.svgNine-layer enterprise reference architecture — Visio-exportable SVG including deployment mapping and reachability layer
How to cite
Priyadarshini, B. (2026). Proven Exploitable: The Practitioner's Guide to AI-Augmented Vulnerability Discovery. Zenodo. https://doi.org/10.5281/zenodo.21159028
Companion paper
Calibrated to Act: The Practitioner's Guide to Building an Agentic SOC That Knows When Not to Act
The companion paper covers the defensive side — how to deploy, govern, and calibrate agentic SOC systems that know when not to act autonomously.
DOI: 10.5281/zenodo.21157411 URL: https://zenodo.org/records/21157411 GitHub: see companion repository
Licence
This paper is published under Creative Commons Attribution 4.0 International (CC BY 4.0).
You are free to share and adapt this paper for any purpose, including commercial use, as long as you give appropriate credit and link back to the original DOI.
Disclosure
Research, drafting, and editorial assistance was provided by Claude (Anthropic). All analysis, decisions, and conclusions are the author's own. All primary sources have been verified against original publications.
About the author
Bindiya Priyadarshini is a security practitioner and researcher specialising in AI-augmented security operations, agentic systems governance, and enterprise security architecture.
Companion paper: https://doi.org/10.5281/zenodo.21157411 Zenodo profile: https://zenodo.org/search?q=Priyadarshini%2C%20Bindiya
The adversaries aren't waiting. Neither should you.