An open-source CLI tool called PentesterFlow has been released. It connects to local or hosted LLMs and assists with the full workflow of penetration testing—from reconnaissance to validation, evidence collection, and reporting—while requiring human approval at each step.
July 7, 2026 · PentesterFlow
An LLM Pentesting Agent That Keeps the Human in the Loop
An open-source, terminal-resident "agentic AI" that connects local or hosted LLMs to real security tooling — recon to reporting — but records a finding only after reproduction is confirmed by request/response evidence and analyst approval.
8
supported model backends — local & hosted
39+
open-source pentesting agents counted in a 2026 survey
~50%
of DVWA solved with a planner + skill dictionaries
5 min
automatic session snapshots; resume anytime
The engagement workflow
Four connected phases, running real tools within a defined scope.
→
Enumeration
surface & assets
→
Exploitation
approval required
→
Reporting
Markdown findings
Tools invoked: Shell/Bash · HTTP · browser capture · Burp Suite bridge. Confirmed findings saved under ./findings/ with PoC, evidence, impact & remediation.
The rule that fights hallucination
A finding is recorded only after reproduction is confirmed by request/response evidence . Operational learning accumulates with no model retraining.
Local knowledge bases
Facts stored as JSONL, deduplicated & secret-redacted, pinned into every system prompt.
Human-in-the-loop
Analyst approval gates every sensitive action — auditable and reproducible.
Where it works today
Handles recon, enumeration & basic exploitation on easier labs
Quantized Qwen models tuned for 16 GB VRAM are viable
Planner + skill dictionaries solved ~half of DVWA
Where caution remains
Smaller / heavily quantized models hallucinate more
Context length constrains long tasks even at 128K
Complex multi-step workflows need scoping & human guidance
Runs broadly: macOS, Linux (arm64/x64) & Windows (x64). x64 builds use Bun's baseline runtime — running even on systems without AVX2 support. Open source and self-hostable.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…