What does Agent-time AppSec actually mean?

MCP stands for Model Context Protocol, the open standard that lets AI coding agents (Claude Code, Cursor, Windsurf, GitHub Copilot, Gemini CLI) call external tools mid-session. CybeDefend registers as an MCP server. When your agent is about to write or modify code, it calls CybeDefend's MCP tool, which runs our reachability-aware SAST, SCA, IaC, secrets and business-logic scanners against that diff in under 1.2 seconds. The agent receives the security verdict, with a fix suggestion, before it proposes the line. This is fundamentally different from post-commit SAST: security runs inside the agent loop, at creation time, not after the PR is opened. No YAML configuration, no separate CI step, no review meeting. The agent enforces your security policy the same way it enforces your code style: automatically, in the prompt, before the first save.

How is this different from a regular SAST scanner?

Legacy SAST runs on commits and floods you with 95% false positives. We run at agent-time on the diff your agent is about to ship, with reachability + framework awareness. Same engine, different surface, 1/10 of the noise.

Which AI agents are supported?

Claude Code, Cursor, Windsurf, GitHub Copilot, Gemini CLI, Cline, Continue.dev and Zed. Plus a generic MCP server for anything else that speaks MCP.

Can we self-host? What about data residency?

Yes. Self-host on the Enterprise plan (single OCI image or air-gapped tar). EU and US regions are fully isolated, you pick where your data lives. CybeDefend SAS is registered in Lille, France; EU workloads are hosted in Paris and US workloads in Iowa.

Find, Fix, Repeat.Secure your

No vuln, no bill.

Book a 30-min demo

Watch the 2-minute platform tour

Claude CodeVibeDefend

Add an endpoint to update user profile

Product tour

See the whole platform in action

AutoFix, Analysis, business-logic research, policies, reports, IDE integrations. The whole platform, in under four minutes.

.platform-video-facade{display:none!important}

Backed by

The reckoning

The problem isn’t that AI writes code.
It’s that no one can read it anymore.

Your best reviewer reads 200 lines a day. Your agent ships 5,000. You don’t need better review. You need security at creation time.

70%+

of new code at AI-native companies is written by AI

Anysphere · Anthropic · public disclosures · 2025-2026

43%

of API vulnerabilities exploit logic, not CVEs

Wallarm · 2026 API ThreatStats

100×

to fix in production vs. at the prompt

Industry consensus · NIST 02-3

Why now

Three years. Three paradigms. Your AppSec stack stayed in 2023.

Each era grew the pile of unread code tenfold. The tooling never moved.

Step 01 - Suggest era

2023

Copilot whispers. SAST scans nightly.

Devs accept suggestions faster than reviewers can read them. AppSec is scanning code no human wrote. Slow, but still auditable.

AI output: lines / day
Review: human
Gate: nightly CI

Step 02 - Act era

2025

Agents drive. Reviewers lag.

Cursor, Claude Code, Windsurf write whole features. SAST still bolted onto CI. Logic flaws land faster than anyone can flag them.

AI output: 1k+ LOC / day
Review: tooling lag
Gate: post-merge CI

Step 03 - Debt era

2026

you are here

Agents ship solo. You inherit the debt.

5,000 LOC / day across services. Logic flaws compound. The pipeline is the bottleneck and the blind spot. The backlog grows faster than any team can clear it. You are here.

AI output: 5k LOC / day
Review: Agent-time
Gate: in the prompt

The foundation

Scanners chase patterns.
We mapped your business
into a graph.

paradigm reversedScannersPattern-matchers. Reactive. Run after the fact.VibeDefendKnowledge graph. Pre-built. Queried before the line.

VibeDefend indexes your codebase once into a Security Code Knowledge Graph: files, functions, tenants, payment paths, the rules you already enforce. The agent queries the map instead of grepping the diff.

Scanners match strings after the fact. VibeDefend answers before the line is suggested.

cybe agent · computing graph…

service-core service

payouts.ts

agent note

Service module. Owns the business rules for moving money: validation, ledger update, risk side-effects. The graph routes most queries through here, touching it triggers the strictest rule pack.

transactional writesrisk hook requiredno shortcut paths

called by

calls

files: 156
functions: 98
edges: 218
rules: 42

Shift-left, by default

Vibe-coding rewrote the rules.
Not the pull request.

Agent, IDE, pipeline: wherever your team ships code, VibeDefend is already inside.

01 · Agents10 native

AI coding agents

VibeDefend runs as an MCP server inside the agent loop. Verdict before the line is suggested.

Claude Code - Cursor - Windsurf - GitHub Copilot - OpenAI Codex - Google Gemini - Cline - Continue - Zed - Google Antigravity

02 · Editors5 editors

IDEs & editors

VS Code, Cursor, Windsurf, JetBrains, Antigravity. Inline diff review, sub-100 ms verdict.

Visual Studio Code - Cursor - Windsurf - Google Antigravity - JetBrains

03 · Pipelines9 engines

CI / CD & pipelines

SARIF-native in every CI. Block the PR, sign the patch, audit the override.

GitHub - GitLab - Azure Pipelines - Jenkins - Azure DevOps - Bitbucket - CircleCI - TeamCity - Atlassian Bamboo

What scanners never see

Eight flaws only the agent can catch. Killed before the first save.

43% of API vulnerabilities exploit business logic, not CVEs (Wallarm, 2026). These eight never trip a scanner. VibeDefend rewrites them in the prompt.

Click any card to see the diff

Autopilot

Zero YAML. Your rules, mined from your own code.

“But who writes all these rules?” We do. Autopilot reads your graph and drafts the rules you already follow as candidates. No security engineer required.

01
Walk the graph
It walks your imports, calls, tenants and routes, every boundary you already enforce.
02
Draft candidate rules
Patterns surface as YAML-free rules with name, scope, severity, and code references.
03
Accept · edit · reject
You stay in control. Review per rule, batch-accept the ones you trust, ignore the rest.

CybeDefend Autopilot - Rule candidatesAutopilotscanning - 3 candidates3

Proposed rule01-Multi-tenancy

Every query on db.orders filters by tenantId

Found in 42 of 43 occurrences across 16 files - confidence 94% - category multi-tenancy

Gap detected-src/handlers/export.ts:17 doesn't comply

Accept ruleEdit scopeReject

Proposed rule02-Authorisation ceiling

Refunds over €500 require role finance.manager

Found in 7 of 8 occurrences across 4 files - confidence 87% - category SoD - authorisation

Gap detected-src/handlers/refund.ts:22 doesn't comply

Accept ruleEdit scopeReject

Proposed rule03-Audit logging

Every payment mutation writes audit.log(actor, amount)

Found in 31 of 34 occurrences across 11 files - confidence 91% - category compliance - audit trail

Gap detected-src/handlers/payouts.ts:54, +2 more

Accept ruleEdit scopeReject

The shift

The 2010s AppSec stack was never built for a world where AI writes half the code.

Line for line: what the legacy stack does, and where one agent-time layer replaces it.

2010s · The legacy stack

Built for a world where humans wrote the code.

Reads syntax (Checkmarx, Sonar). Blind to intent, blind to business logic.
Up to 99% false-positive rate. 21,000 engineering hours burned per year.
Per-seat pricing (up to $15k/user/year at Veracode). Growth-taxed.
Blind to AI agents. Nothing between the prompt and the PR.
Runs in CI, after the fact. Devs see findings hours later.
Six or seven vendors, six or seven dashboards, six or seven invoices.
Generic advice. No idea what your codebase actually does.

2026 · The CybeDefend way

Agent-time AppSec. One layer. In the prompt.

Reads intent. Tenant leaks, refund bypass, PII, idempotency, SoD violations, the logic flaws syntax scanners miss.
95% of noise filtered. Reachability scoring on the graph, not blind line-by-line.
One flat subscription. Not per seat, not per repo. Whole company inside.
Every engine in one graph. AI-BOM · SAST · SCA · Secrets · License · IaC · Container · CI/CD · Business-Logic.
Lives inside every AI agent. Claude Code, Cursor, Copilot, one semantic layer across all.
Runs in the IDE. Cleared before the PR is opened.
Knows your codebase. Your rules, your tenants, your business, encoded in the graph.

ROI Calculator

Your team. Your code. Your ROI in 30 seconds.

Five questions. See the hours your team gets back, the triage budget you stop burning, and the net annual gain on every plan.

ROI Calculator

Voices

Voices from teams already shipping with VibeDefend in the loop.

Two early customers, two different stacks, same outcome: more time shipping, fewer alerts to triage.

Olivier, Tech Lead - Koddex — OL
Olivier
Tech Lead

Geoffrey, CTO - Diag n’Grow — GE
Geoffrey
CTO

Live · just shipped

Install VibeDefend in 5 seconds.

One command wires every coding agent on your machine to CybeDefend: your business rules, your compliance frameworks, and guards that block destructive calls before they fire.

Install in 5 secondsNode 18.17+

npx -y @cybedefend/vibedefend@latest install

Auto-detects

Claude Code
Cursor
OpenAI Codex
Windsurf
VS Code Copilot

Read the README on npm

Find, Fix, Repeat.Secure your

See the whole platform in action

Backed by

The problem isn’t that AI writes code.
It’s that no one can read it anymore.