Daily Reading for 2026-09-16

AI-Assisted Threat Modeling

A developer’s reflection on Shostack & Associates’ first four-day Threat Modeling Intensive with Complete AI at Black Hat, pairing the classic Four Question Framework with hands-on LLM use. The throughline is that bringing AI into threat modeling is still fundamentally a people and practice problem: attendees range from never having used an LLM at work to using one daily, yet nearly everyone reports some form of confusion about how, when, and why AI fits a threat-modeling workflow. It matters because threat modeling is shifting from a craft practiced mainly by security specialists into an AI-augmented skill that developers are increasingly expected to adopt in ordinary agile and product loops.

Watch for:

• The OWASP Global AppSec USA (Nov 2–4, 2026) course pairing threat modeling with LLM hands-on experience
• Whether AI-assisted threat-modeling methods become standardized enough to spread beyond Shostack’s training
• How the Four Question Framework absorbs AI tooling without losing its discipline

Wardley signals:

• Components: The Four Question Framework, STRIDE, data-flow diagrams, LLM assistants for threat modeling, Shostack + Associates training courses

• Evolution stage: Threat modeling itself is Product (established courses and practice); the LLM-assisted layer is Genesis-to-Custom — practice is novel, unstandardized, and still being figured out by trainers and students alike

• Doctrine: “Use appropriate tools” (apply AI where it helps), “Know your users” (developers arrive differently than security experts), and “Challenge Assumptions” (the near-universal LLM confusion)

• Climatic patterns: Co-evolution of practice with activity — AI tooling is evolving in lockstep with how threat modeling is taught and performed

• Diffusion markers: The jump from two-day to four-day courses, new dedicated LLM-for-threat-modeling offerings, and mounting conference-circuit uptake

• Threat Modeling Intensive with Complete AI: From a Developer’s POV

wardley-beta
%% 1. Users: developers and security engineers learning threat modeling
%% 2. Needs: know how, when, and why to use AI in threat-modeling workflows
%% 3. Capabilities: threat-modeling education, structured threat brainstorming, AI-assisted analysis
%% 4. Components: Four Question Framework, STRIDE brainstorming, data-flow diagrams, LLM threat-modeling assistants, training courses
%% 5. Maturity: LLM assist (Genesis→Custom), Four Question Framework (Product), data-flow diagrams (Commodity), training courses (Product)
%% 6. Managed by: Shostack & Associates trainers and attendees

anchor developerNeed [0.95, 0.92]
anchor securityNeed [0.92, 0.86]

component fourQuestionFramework [0.55, 0.75]
component strideBrainstorming [0.45, 0.70]
component dataFlowDiagrams [0.72, 0.62]
component llmThreatAssist [0.20, 0.55]
component trainingCourse [0.40, 0.82]

fourQuestionFramework -> developerNeed
trainingCourse -> developerNeed
fourQuestionFramework -> securityNeed
strideBrainstorming -> fourQuestionFramework
dataFlowDiagrams -> fourQuestionFramework
llmThreatAssist -> fourQuestionFramework

evolve llmThreatAssist 0.30
evolve fourQuestionFramework 0.60
evolve strideBrainstorming 0.50
evolve dataFlowDiagrams 0.80
evolve trainingCourse 0.52

Red Heron Gitea Exploitation

Acronis TRU documents Red Heron, a PRC-linked actor that weaponized the critical Gitea diffpatch RCE (CVE-2026-60004, CVSS 9.8) within days of disclosure, scanning 1,386 exposed Gitea instances across seven countries and confirming compromises in Canada, Argentina, Taiwan, the US, and Sri Lanka. The operation escalated from source-code theft to credential collection, SSH persistence, and lateral movement, including root access across a three-node Proxmox cluster. A rare exposed staging server revealed the full toolkit: the JITTERLY C++ implant and the previously undocumented SIXZUT LD_PRELOAD rootkit. It matters because self-hosted development platforms are a fast, high-value attack surface where n-days turn into automated intrusion in days.

Watch for:

• Patch adoption of Gitea 1.27.1 versus lingering exposure across 1.17–1.27.0 instances
• Detections of the SIXZUT rootkit and JITTERLY implant, plus its Adaptix C2 lineage
• Whether Red Heron’s automation spreads to other self-hosted dev platforms or campaigns

Wardley signals:

• Components: Gitea, CVE-2026-60004 diffpatch RCE, the JITTERLY implant, the SIXZUT LD_PRELOAD rootkit, Adaptix C2, Proxmox

• Evolution stage: Self-hosted source-code management is Commodity (widely deployed, accepted); the actor’s exploitation and persistence tooling is Genesis-to-Custom — bespoke, per-campaign automation rather than a standard offering

• Doctrine: Defenders miss “Know the details” and “Set exceptional standards” by leaving vulnerable instances exposed; the actor models “Focus on the outcome” by moving swiftly from PoC to automated weapon

• Climatic patterns: The Red Queen — attackers automate faster than defenders patch; “not everything survives” as high-value instances fall quickly

• Diffusion markers: The rapid PoC-to-automated-weaponization cycle and copy-cat n-day campaigns emerging across self-hosted platforms

• 🚩 Red Heron Exploits Gitea CVE-2026-60004 to Deploy JITTERLY and SIXZUT Linux Rootkit

wardley-beta
%% 1. Users: organizations running self-hosted Gitea and their defenders
%% 2. Needs: keep source code and credentials safe, patch n-days before intrusion
%% 3. Capabilities: patching, vulnerability scanning, detection of implants and rootkits
%% 4. Components: Gitea, CVE-2026-60004 diffpatch RCE, JITTERLY implant, SIXZUT rootkit, patch management, vulnerability scanning
%% 5. Maturity: Gitea (Commodity), diffpatch RCE (Genesis→Custom), JITTERLY/SIXZUT (Genesis), patch mgmt (Product), vuln scanning (Product)
%% 6. Managed by: operators and security vendors; weaponized by Red Heron

anchor orgDefenders [0.95, 0.90]

component giteaPlatform [0.75, 0.70]
component diffpatchRce [0.15, 0.55]
component jitterlyImplant [0.15, 0.50]
component sixzutRootkit [0.10, 0.45]
component patchManagement [0.45, 0.65]
component vulnScanning [0.50, 0.60]

giteaPlatform -> orgDefenders
patchManagement -> giteaPlatform
vulnScanning -> patchManagement
diffpatchRce -> giteaPlatform
jitterlyImplant -> giteaPlatform
sixzutRootkit -> jitterlyImplant

evolve giteaPlatform 0.80
evolve diffpatchRce 0.25
evolve jitterlyImplant 0.20
evolve patchManagement 0.60
evolve vulnScanning 0.65

Cloud Identity Behavioral Clustering

Unit 42 applies unsupervised machine learning (UMAP + HDBSCAN) to AWS CloudTrail logs, clustering more than 40,000 identities across 125 cloud environments into functional roles — administrators, backup services, security tooling, and DevOps — based on observed behavior rather than resource names or IAM policies. The aim is to catch role masquerading, where attackers exploit pre-existing permission profiles and benign labels to hide malicious activity. The research also distills the clustering map into lightweight SQL heuristics so teams can classify identities at scale without continuously running an ML pipeline. It matters because over-privileged identities make the distinction between what an identity can do and what it actually does the central detection question.

Watch for:

• The SQL heuristic logic migrating into real detection rules and query engines
• Whether the behavioral-clustering method generalizes to SaaS and Kubernetes audit logs as claimed
• Cloud detection and response (CDR) products adopting behavior-based role baselines

Wardley signals:

• Components: AWS CloudTrail, UMAP, HDBSCAN, IAM, CSPM, CDR, and Palo Alto’s Cortex Cloud / XSIAM / Idira products

• Evolution stage: Cloud identity security is Product-to-Commodity (IAM and CSPM are accepted practice); behavioral role mapping is Genesis-to-Custom — research-grade ML not yet standard tooling, being simplified into SQL

• Doctrine: “Know your users” (behavior over permissions), “Bias towards data” (audit-log-driven), and “Use appropriate tools” (unsupervised ML followed by lightweight heuristics)

• Climatic patterns: Co-evolution — attacker masquerading practice evolves alongside defenders’ behavioral baselines

• Diffusion markers: The methodology moving from research into product, and its stated extension beyond AWS to other providers, SaaS, and Kubernetes

• Unit 42 Maps Cloud Identity Behavior to Detect Role Masquerading and Anomalous Activity

wardley-beta
%% 1. Users: cloud security teams
%% 2. Needs: detect role masquerading and anomalous identity behavior
%% 3. Capabilities: identity behavior mapping, role classification, automated detection
%% 4. Components: AWS CloudTrail, IAM, behavioral clustering (UMAP+HDBSCAN), SQL heuristics, CDR detection, CSPM audit
%% 5. Maturity: CloudTrail/IAM (Commodity), behavioral clustering (Custom), SQL heuristics (Custom→Product), CDR (Product), CSPM (Product)
%% 6. Managed by: Unit 42 / Palo Alto and cloud security teams

anchor cloudSecurityTeam [0.95, 0.90]

component cloudtrailLogs [0.85, 0.55]
component iamPolicies [0.80, 0.60]
component behavioralClustering [0.25, 0.65]
component sqlHeuristics [0.45, 0.70]
component roleMasqueradeDetection [0.55, 0.78]
component cspmAudit [0.70, 0.68]

roleMasqueradeDetection -> cloudSecurityTeam
cspmAudit -> cloudSecurityTeam
sqlHeuristics -> roleMasqueradeDetection
behavioralClustering -> sqlHeuristics
cloudtrailLogs -> behavioralClustering
iamPolicies -> roleMasqueradeDetection

evolve behavioralClustering 0.30
evolve sqlHeuristics 0.55
evolve roleMasqueradeDetection 0.60
evolve cspmAudit 0.75
evolve cloudtrailLogs 0.88

FortiGate VPN Exploitation Campaign

Hunt.io’s AttackCapture discovered an open staging directory exposing an active intrusion against 3BB (Triple T Broadband) in Thailand. The operator exploited CVE-2024-21762 on a FortiGate 60F SSL-VPN, installed MeshCentral as a persistent backdoor (device group TH-3BB, C2 at ayuthayatech.com), harvested RADIUS subscriber credentials, and pre-staged a cleanup script to erase logs while preserving the agent. The 298-file arsenal spanned PwnKit and Dirty COW privilege escalation, SSH brute-forcing, Ghostcat, F5 exploits, and a valid OpenVPN certificate minted from the victim’s own PKI. It matters as a rare open-window view of a VPN-to-internal-network compromise and why SSL-VPN edge devices stay prime targets.

Watch for:

• FortiGate / FortiOS patch coverage for CVE-2024-21762 across exposed edge devices
• MeshCentral-abuse detections, since legitimate remote-management tooling is used as C2
• RADIUS and subscriber credential dumps resurfacing against other telecom providers

Wardley signals:

• Components: FortiGate 60F / FortiOS, CVE-2024-21762, MeshCentral, RADIUS, OpenVPN, PwnKit, Dirty COW, Hunt.io AttackCapture

• Evolution stage: SSL-VPN is Commodity edge infrastructure (widely deployed and accepted); the MeshCentral-as-C2 persistence pattern is Custom — creative abuse of an off-the-shelf product rather than a standard practice

• Doctrine: Operators miss “Manage failure” and “Set exceptional standards” by running unpatched edge VPNs; the actor applies “Focus on the outcome” and “Know the details” of the victim’s environment

• Climatic patterns: “Success breeds inertia” — persistent legacy VPN infrastructure invites attack; componentisation lowers cost as cheap staging hosts enable intrusion

• Diffusion markers: Recurring MeshCentral persistence patterns across ISPs and open-directory visibility accelerating threat-intelligence sharing

• 🚩 FortiGate Exploitation Against Thai Broadband Provider Leads to MeshCentral Persistence, Credential Theft, and Internal Lateral Movement

wardley-beta
%% 1. Users: telecom and enterprise network defenders
%% 2. Needs: secure edge VPNs, detect persistence and credential theft
%% 3. Capabilities: patching, intrusion detection, persistent-access visibility
%% 4. Components: FortiGate/FortiOS SSL-VPN, CVE-2024-21762, MeshCentral backdoor, RADIUS credential store, edge patch mgmt, intrusion detection
%% 5. Maturity: SSL-VPN (Commodity), CVE exploit (Genesis→Custom), MeshCentral abuse (Custom), RADIUS (Commodity), patch mgmt (Product), intrusion detection (Product)
%% 6. Managed by: network operators; exploited by a Thai-based actor

anchor networkDefenders [0.95, 0.90]

component fortigateSslvpn [0.78, 0.65]
component cve21762Exploit [0.15, 0.55]
component meshcentralBackdoor [0.30, 0.60]
component radiusCredStore [0.70, 0.60]
component edgePatchMgmt [0.45, 0.70]
component intrusionDetection [0.55, 0.75]

fortigateSslvpn -> networkDefenders
intrusionDetection -> networkDefenders
cve21762Exploit -> fortigateSslvpn
meshcentralBackdoor -> fortigateSslvpn
radiusCredStore -> fortigateSslvpn
edgePatchMgmt -> fortigateSslvpn

evolve fortigateSslvpn 0.80
evolve cve21762Exploit 0.25
evolve meshcentralBackdoor 0.35
evolve radiusCredStore 0.75
evolve edgePatchMgmt 0.60
evolve intrusionDetection 0.62

Datacenter Buildout Moratoriums

SemiAnalysis pushes back on the narrative that local moratoriums are killing the US datacenter buildout. Mapping more than 300 moratoriums across towns, cities, and counties plus state actions (New York’s permit pause, Texas’s ERCOT queue pause, Pennsylvania, Oregon), they estimate only about 2.3GW of planned capacity is genuinely delayed by policy — against a forecast of +38GW of US datacenter IT capacity delivered in 2027, more than double 2026. Roughly 20GW sits inside restricted boundaries, but most projects proceed anyway. Their model tracks 6,000+ datacenters via satellite imagery. It matters because hype-driven policy narratives distort capacity planning and energy decisions.

Watch for:

• The real delivery impact of New York’s executive-order permit pause
• Movement in the ERCOT interconnection queue in Texas
• Whether satellite-tracked construction confirms the +38GW 2027 forecast

Wardley signals:

• Components: Datacenters, the SemiAnalysis datacenter industry model, satellite-imagery tracking, ERCOT interconnection queue, state and local permitting, electricity and water infrastructure

• Evolution stage: Datacenter capacity is Commodity-to-Product (rapid industrialization driven by AI demand); permitting and policy are Custom (high local variation and uncertainty)

• Doctrine: “Bias towards data” (satellite-tracked evidence over narrative) and “Challenge Assumptions” (debunking the moratorium myth)

• Climatic patterns: Jevons’ paradox / efficiency-enabled innovation — AI demand keeps accelerating despite constraints; a regulatory “economic state of war” reacting to a boom

• Diffusion markers: Policy copy-cat adoption across 300+ jurisdictions and a public war of forecasts between capacity models

• Everyone Says Datacenter Moratoriums Are Killing the US Buildout. We disagree

wardley-beta
%% 1. Users: datacenter operators and capacity planners
%% 2. Needs: accurate buildout picture and reliable capacity planning
%% 3. Capabilities: forecasting, construction tracking, policy-aware planning
%% 4. Components: datacenter capacity, satellite tracking, industry forecast model, permitting policy, ERCOT interconnect, grid power supply
%% 5. Maturity: datacenter capacity (Product→Commodity), satellite tracking (Custom), forecast model (Product), permitting (Custom), ERCOT (Commodity), grid (Commodity)
%% 6. Managed by: operators, utilities, and state/local regulators

anchor operatorsNeed [0.95, 0.90]

component datacenterCapacity [0.75, 0.70]
component satelliteTracking [0.35, 0.60]
component industryForecastModel [0.45, 0.68]
component permittingPolicy [0.30, 0.72]
component ercotInterconnect [0.70, 0.55]
component gridPowerSupply [0.80, 0.50]

datacenterCapacity -> operatorsNeed
industryForecastModel -> datacenterCapacity
satelliteTracking -> industryForecastModel
permittingPolicy -> datacenterCapacity
ercotInterconnect -> datacenterCapacity
gridPowerSupply -> datacenterCapacity

evolve datacenterCapacity 0.78
evolve satelliteTracking 0.40
evolve industryForecastModel 0.50
evolve permittingPolicy 0.35
evolve ercotInterconnect 0.75
evolve gridPowerSupply 0.85

Agent Orchestration and Token Economics

Steve Yegge’s 4am dump on the “Fuel Apocalypse”: his Wheelhouse orchestrator (a Beads machine in the Gas Town lineage) runs about 25 Claude Fable instances plus 25 Opus and Sol agents, pushing 250–300 meaningful commits a day on his game Wyvern — but token costs are compounding faster than output. Fable 5.1 pricing keeps rising; he burned through 21 Claude Max accounts and stopped, estimating that sustaining 24x7 at current pricing would cost around $12,000/month. His catalogue of fuel-efficiency levers — routing to cheaper models, earlier handoff, smaller tasks, better startup context, prompt caching, off-peak runs — frames token economics as the binding constraint on real-world AI factories.

Watch for:

• Fable-tier model pricing and availability over the next 12 months
• Fuel-efficiency techniques (routing, caching, handoff, off-peak) becoming shared best practice
• Whether orchestrators like Wheelhouse and Gas Town commoditize as frameworks mature

Wardley signals:

• Components: Claude Fable 5 / 5.1, Wheelhouse, Gas Town, Beads machines, Claude Max accounts, token pricing, prompt caching

• Evolution stage: Frontier-model intelligence is Product (high unit cost, still proprietary); agent/orchestrator frameworks are Genesis-to-Custom (novel and bespoke); token “fuel” economics is the current bottleneck

• Doctrine: “Do better with less” and “Effectiveness over efficiency” — the entire post is about fuel efficiency; “Bias towards the new” (agents) and “There is no core” (orchestrators as infrastructure)

• Climatic patterns: Jevons’ paradox — more capable models consume ever more tokens; efficiency enabling innovation; creative destruction as pricing shifts reshape the landscape

• Diffusion markers: Rapid agent-framework iteration (Gas Town → Wheelhouse), repeated frontier pricing resets, and a community pivot toward fuel-efficiency playbooks

• Seats and Sunsets

wardley-beta
%% 1. Users: developers running AI agent factories / orchestrators
%% 2. Needs: deliver real work affordably while managing token ("fuel") cost
%% 3. Capabilities: agent orchestration, token cost control, quality routing
%% 4. Components: frontier model, orchestrator framework, agent workers, token pricing, prompt caching, fuel-efficiency playbook
%% 5. Maturity: frontier model (Product), orchestrator (Custom), agent workers (Product), token pricing (Product), prompt caching (Product→Commodity), fuel playbook (Custom)
%% 6. Managed by: orchestrator owners (e.g. Yegge's Wheelhouse) and model vendors

anchor aiFactoryOwner [0.95, 0.90]

component orchestratorFramework [0.20, 0.60]
component agentWorkers [0.40, 0.72]
component frontierModel [0.50, 0.65]
component tokenPricing [0.35, 0.50]
component promptCaching [0.55, 0.45]
component fuelEfficiencyPlaybook [0.25, 0.78]

orchestratorFramework -> aiFactoryOwner
fuelEfficiencyPlaybook -> aiFactoryOwner
agentWorkers -> orchestratorFramework
frontierModel -> agentWorkers
tokenPricing -> orchestratorFramework
promptCaching -> frontierModel
fuelEfficiencyPlaybook -> tokenPricing

evolve orchestratorFramework 0.30
evolve agentWorkers 0.50
evolve frontierModel 0.55
evolve tokenPricing 0.45
evolve promptCaching 0.65
evolve fuelEfficiencyPlaybook 0.35

Gemini Live Speech Models

Simon Willison covers Google’s release of Gemini 3.8 Live and 3.8 Live Extended Thinking, two speech-to-speech models in the same shape as OpenAI’s GPT-Live family. He had GPT-6 Astra Extra High build a browser UI for the new models against the WebSocket BidiGenerateContent endpoint, using the Web Audio API for capture and playback, no libraries, and support for live interruption. It matters as a hands-on pointer to the current state of real-time voice interaction APIs and how quickly new speech models are being shipped and consumed.

Watch for:

• Gemini Live Extended Thinking’s output quality versus the GPT-Live family
• Stabilization and standardization of the WebSocket voice interaction API
• Third-party voice UIs and apps built on the new Gemini Live models

Wardley signals:

• Components: Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking, GPT-Live / GPT-6 Astra, the BidiGenerateContent WebSocket API, the Web Audio API

• Evolution stage: Speech-to-speech models are Product — emerging, rapidly iterating, with competing forms; the WebSocket and Web Audio plumbing underneath is Commodity infrastructure

• Doctrine: “Bias towards the new” and “Use appropriate tools” (a no-libraries WebSocket UI); “Exploit the landscape” by building directly on the new API

• Climatic patterns: Componentisation — new voice models build on commodity WebSocket and audio infrastructure, speeding recombination and evolution

• Diffusion markers: Google cloning the GPT-Live shape signals competitive convergence; tutorial and API uptake spreading real-time voice interaction patterns

• Gemini Live audio

wardley-beta
%% 1. Users: developers building real-time voice apps and their end users
%% 2. Needs: natural, interruptible speech-to-speech interaction
%% 3. Capabilities: speech synthesis/understanding, streaming transport, live interruption
%% 4. Components: Gemini 3.8 Live, GPT-Live family, WebSocket voice API, Web Audio API, voice UI, live interruption
%% 5. Maturity: speech-to-speech models (Product), WebSocket voice API (Product→Commodity), Web Audio (Commodity), voice UI (Custom)
%% 6. Managed by: Google/OpenAI and app developers

anchor voiceAppUsers [0.95, 0.90]

component geminiLiveModel [0.55, 0.70]
component gptLiveFamily [0.55, 0.66]
component websocketVoiceApi [0.70, 0.55]
component webAudioCapture [0.80, 0.50]
component voiceUi [0.35, 0.78]
component interruptionSupport [0.45, 0.74]

voiceUi -> voiceAppUsers
geminiLiveModel -> voiceUi
gptLiveFamily -> voiceUi
websocketVoiceApi -> geminiLiveModel
webAudioCapture -> voiceUi
interruptionSupport -> voiceUi

evolve geminiLiveModel 0.60
evolve websocketVoiceApi 0.70
evolve webAudioCapture 0.85
evolve voiceUi 0.45
evolve interruptionSupport 0.55

AI-Driven Security Research Roundup

YesWeHack’s roundup centers on James Kettle’s “autonomous research cascade” — an AI system that invented new HTTP desync techniques and compromised banks, security solutions, and government infrastructure, presented at Black Hat and DEF CON. It also covers CRLF-powered desync worms (Stacey & Righi), CSS webmail exploitation (Gareth Heyes), and UniBLEed unauthenticated root RCE on Unitree G1 humanoid robots, plus a spotlight on silent patch gaps, Codex and Claude Code in bug-bounty pipelines, and the YesWeHack Claude Kit. The throughline is AI radically lowering the cost of both finding and exploiting vulnerabilities, widening the defender-attacker asymmetry.

Watch for:

• Release of Kettle’s HTTP Terminator / autonomous research cascade tooling
• Silent-patch-gap advisories (e.g. Proxmox CVE-2023-54391) catching up with exploitation
• Agentic coding tools (Codex, Claude Code) moving into real bug-bounty pipelines

Wardley signals:

• Components: HTTP Terminator, CRLF desync techniques, CSS/webmail attacks, UniBLEed (Unitree G1 BLE), Codex, Claude Code, Burp Suite, the YesWeHack Claude Kit

• Evolution stage: AI-assisted vulnerability research is Genesis-to-Custom (novel, researcher-led, no standard method); HTTP desync technique classes are Custom research; robot/device hacking is Genesis

• Doctrine: “Bias towards data” (evidence-grounded reports) and “Challenge Assumptions”; researchers model “Focus on the outcome” while defenders lag

• Climatic patterns: The Red Queen — AI accelerates attacker rediscovery of silent patches; creative destruction of vulnerability-intelligence norms

• Diffusion markers: Research cascading into tooling and commercial platforms, agentic-tool adoption in bug bounty, and a “100% human-found” counter-narrative pushing back

• James Kettle’s ‘autonomous research cascade’, CRLF-powered desync attacks, RCE on humanoid robots – ethical hacker news roundup

wardley-beta
%% 1. Users: security researchers, bug bounty hunters, and defenders
%% 2. Needs: discover vulnerabilities faster and keep up with AI-assisted research
%% 3. Capabilities: AI-driven research, desync technique development, agentic exploitation, evidence-grounded reporting
%% 4. Components: AI-assisted vuln research, HTTP desync techniques, agentic coding tools, robot RCE research, bug bounty platform, silent patch intel
%% 5. Maturity: AI-assisted research (Genesis→Custom), desync techniques (Custom), agentic tools (Product), robot RCE (Genesis), bug bounty (Product), silent patch intel (Custom)
%% 6. Managed by: researchers (Kettle, PortSwigger), YesWeHack, and vulnerability intel providers

anchor securityResearchers [0.95, 0.90]
anchor defenders [0.90, 0.84]

component aiVulnResearch [0.20, 0.70]
component httpDesyncTechniques [0.25, 0.62]
component agenticCodingTools [0.45, 0.68]
component robotRceResearch [0.08, 0.55]
component bugBountyPlatform [0.60, 0.72]
component silentPatchIntel [0.30, 0.58]

aiVulnResearch -> securityResearchers
robotRceResearch -> securityResearchers
bugBountyPlatform -> securityResearchers
silentPatchIntel -> defenders
httpDesyncTechniques -> aiVulnResearch
agenticCodingTools -> aiVulnResearch
silentPatchIntel -> securityResearchers

evolve aiVulnResearch 0.30
evolve httpDesyncTechniques 0.35
evolve agenticCodingTools 0.50
evolve robotRceResearch 0.15
evolve bugBountyPlatform 0.65
evolve silentPatchIntel 0.40