wakeupdexter 1.0.2

"He doesn't just find flaws. He verifies them."

Local-first, agentic application security assessment — that investigates its own findings before it trusts them.

Dexter doesn't run a fixed set of checks and dump a static report. It runs every scan through a closed loop — Scan → Analyze → Verify → Refine — that investigates its own findings before it trusts them.

dexter --target ./your-project -n --instruction "prioritize secrets and access control"

---

Why Dexter?

Most vulnerability scanners share the same weaknesses: high false-positive rates, no correlation between tools, and no transparent path from "flagged" to "confirmed." Dexter's Verify stage is a real tool-calling agent (ReAct pattern) — it reads surrounding code, computes entropy on suspected secrets, and greps the codebase for related usage before handing back a plain-English verdict instead of just a severity label.

Every tool can also run sandboxed: a disposable, per-scan Docker container instead of your host machine. See Docker Sandbox.

The SAVR Loop

Stage	What happens
Scan	14 integrated tools + local static-analysis rules
Analyze	Base confidence score assigned per finding
Verify	Agentic investigation — reads code, checks entropy, searches the repo
Refine	Instruction-aware confidence boost, bounded by a circuit breaker

Both entry points — the one-shot dexter CLI (CI-friendly, non-zero exit on findings) and the interactive wakeupdexter REPL — call the same underlying engine. No logic is duplicated between them.

---

Quickstart

python -m venv .venv
.venv\Scripts\activate        # Windows — use `source .venv/bin/activate` on macOS/Linux
pip install wakeupdexter

dexter --target ./your-project -n --instruction "prioritize secrets and access control"
wakeupdexter                  # or launch the interactive command center

Recommended: build the sandbox once so tools don't need host installation.

dexter sandbox build
set DEXTER_SANDBOX=1          # Windows CMD — PowerShell: $env:DEXTER_SANDBOX="1"
dexter --target ./your-project -n

---

Docker Sandbox

With DEXTER_SANDBOX=1 set, every containerized tool runs inside a disposable, per-run container. The target is bind-mounted read-only at /workspace, and the container is destroyed the moment the scan finishes. No per-platform tool installs, no PATH issues, no Windows-specific tool ports.

dexter sandbox build                  # fast path (~5–10 min)
dexter sandbox build --with-browser   # + Chromium/agent-browser (~20–30 min, not yet used by any adapter)
dexter sandbox status                 # check Docker + image status

If Docker isn't available, Dexter prints a notice and falls back to running tools on the host — DEXTER_SANDBOX=1 is always safe to leave set.

Scanning something running on your own machine? Containers don't share the host's localhost. Use Docker Desktop's special DNS name instead:

dexter authorize http://host.docker.internal:5000
dexter --target http://host.docker.internal:5000 -n

---

Integrated Tools

14 tools total, spanning SAST, secrets, dependency CVEs, and authorized recon/DAST.

Tool	Type	Gate	Sandboxed
Semgrep	SAST	none	✅
Bandit	Python SAST	none	✅
Gitleaks	Secrets	none	✅
Trufflehog	Secrets + live-credential check	none	✅
Trivy	Dependency CVEs	none	✅
Retire.js	Frontend dependency CVEs	none	✅
ESLint (security ruleset)	JS/TS SAST	none	✅
ast-grep	Structural SAST	none	✅
httpx	Recon / fingerprint	authorization	✅
katana	Web crawler	authorization	✅
ffuf	Content discovery	authorization	host-only
Nmap	Port/service recon	authorization	host-only
Nuclei	Template-based DAST	authorization	host-only
Nikto	Web misconfig	authorization	host-only*
OWASP ZAP	DAST	authorization + running daemon	host-only
sqlmap	Exploitation	authorization + exploit command	host-only
jwt_tool	JWT weakness testing	exploit command	host-only
Metasploit	Exploitation	—	not auto-wired, by design

* Perl-based; needs Strawberry Perl on Windows.

Full sandbox parity for the host-only tools is tracked in Roadmap.

---

Safety & Authorization

Dexter treats "authorized targets only" as enforced behavior, not a suggestion:

• Live-target tools refuse to run against anything not explicitly added via dexter authorize <target>.
• Exploitation-tier tools (sqlmap, jwt_tool) are structurally outside the automatic scan registry — enforced by a dedicated test, not just convention.
• Exploit-tier actions require an explicit, separate command and a confirmation prompt in the interactive dashboard — never triggered by a normal scan.
• Metasploit has no automatic invocation at all.
• Even authorized, sandboxed scans mount the target read-only.

Only test systems you own or have explicit written permission to assess.

---

Usage

dexter --target ./app --scan-mode standard
dexter -n -t ./app --instruction "Prioritize authentication and access control"
dexter --target-list ./targets.txt --run-name ci-scan

dexter report ci-scan --format human
dexter report ci-scan --format markdown --output report.md
dexter view ci-scan

dexter tools
dexter auth
dexter authorize <target>

dexter exploit sqlmap "<url>" --param id
dexter exploit jwt_tool "<jwt>"

dexter sandbox build [--with-browser]
dexter sandbox status

wakeupdexter    # interactive REPL

Inside wakeupdexter:

scan <path-or-url> [instructions]              full SAVR loop
/semgrep, /bandit, /gitleaks, /trivy <path>    run one tool directly
/nmap, /nuclei, /nikto, /zap, /ffuf <target>   run one live-target tool (needs authorize first)
exploit sqlmap <url> --param <name>            explicit exploitation step, asks for confirmation
authorize <target>                             allow a live target for gated tools
/tools                                         show installed/missing tool status
runs · report [run] [human] · view [run] · pool

Exit status is non-zero when findings are present — -n mode is CI-safe.

---

LLM Provider Setup

Create ~/.dexter/.env (Windows: %USERPROFILE%\.dexter\.env):

DEXTER_LLM_KEYS=key-one,key-two,key-three

Variable	Purpose
DEXTER_LLM_KEYS	Comma-separated key pool. A 429 cools that key down and rotates to the next.
DEXTER_LLM_BASE_URL	Any OpenAI-compatible endpoint — Groq by default.
DEXTER_LLM_LOCAL_URL	Fallback tier when every pooled key is cooling — point at local Ollama or llama.cpp.
DEXTER_LLM_COOLDOWN_SECONDS	Cooldown duration per key (default 5 hours).

dexter auth shows live status of every key and the local fallback. Dexter functions fully offline with zero API keys via its deterministic fallback.

Never commit a real key. If one ends up in a scanned project's .env, Dexter's own Gitleaks/Trufflehog adapters will correctly flag it — that's not a false positive, rotate it.

---

Architecture

                     dexter (CLI)        wakeupdexter (interactive REPL)
                          │                        │
                          └────────────┬───────────┘
                                        ▼
                              dexter.loop.run_savr()
                                        │
        ┌───────────────┬──────────────┼──────────────┬───────────────┐
        ▼               ▼              ▼               ▼               ▼
      SCAN           ANALYZE        VERIFY          REFINE          REPORT

containers/Dockerfile     sandbox image (Kali-based)
src/dexter/
├── cli.py                one-shot CLI entry point
├── dashboard.py          wakeupdexter interactive REPL
├── loop.py               the SAVR loop; opens/tears down the sandbox per target
├── agentic_verify.py     ReAct tool-calling verification agent
├── tools.py              external tool adapters + registry (sandbox-aware)
├── sandbox.py            per-run Docker container lifecycle
├── scanner.py            local static-analysis rules
├── provider_pool.py      LLM key rotation + tool-calling support
├── authorization.py      authorized-target list
├── report.py             human-language report generation
├── env_file.py           ~/.dexter/.env loader
└── wordlists/common.txt  bundled ffuf wordlist
tests/                    65 tests, isolated from network/tool/Docker availability

---

Roadmap

• Full sandbox parity for Nmap, Nuclei, Nikto, ffuf, sqlmap, jwt_tool
• Cross-tool finding correlation & deduplication
• Agentic Analyze and Refine stages
• Multi-agent coordinator spawning per-stage subagents
• Historical run diffing across scans
• Browser-driven verification (Chromium/agent-browser already installable via --with-browser)
• Metasploit resource-script passthrough
• Revisit the web dashboard (currently shelved on backup-web-platform)

---

Team

Name	Role
Kartik R. Pagariya
Vikrant K. Kadam
Aditya U. Dengale
Pranali D. Yelavikar
Dr. Parikshit Mahalne	Project Guide

VIT — AI & Data Science, Final Year Project

License

MIT

---

Only test systems you own or have explicit written permission to assess.