Turning attack paths into board-ready remediation plans
Drishti AI scans a network, prices its attack paths in dollars, and drafts remediation playbooks with an LLM — because security findings die in jargon, not in the scanner.
Every network audit I've seen fail the same way is not the scan. The scanner does its job. What fails is the handoff: pentesters find an attack path and hand a 40-page PDF full of CVSS scores and protocol names to people who budget in dollars. The finding is technically correct and commercially inert, so it gets deprioritized until the day it's an incident.
Drishti AI is my answer to that handoff problem. It's a network risk scanner that doesn't stop at "here's the vulnerability" — it visualizes the attack graph, prices the blast radius, and drafts the remediation playbook, so the output reads like a decision memo instead of a technical exhibit. It was the core of a project that later won an innovation final at InnoFusion 3.0.
The three-part pipeline
1. Scan and reachability. FastAPI drives the scan: port discovery plus reachability checks across the target range. This is the boring, well-understood half — the part every tool does.
2. Model the attack graph. Raw open ports are not risk. A port on an internal print server is not a port on the internet-facing payment box. Drishti builds a graph where nodes are assets and edges are exploitable reachability, then computes the paths an attacker would actually walk. The interesting output isn't "host X has 12 vulnerabilities" — it's "an unpatched SMB endpoint reaches your domain controller, and that path costs $Y to exploit."
3. Price it and draft the fix. This is where the LLM earns its place. Groq (LLaMA 3) takes the priced attack paths and drafts remediation playbooks: what to close, what to patch, what to isolate, in what order. The front end renders the graph in an interactive Spline 3D mesh so a non-technical viewer can *see* the path, not just read about it.
Why "price the risk" is the whole point
Security teams are judged on things they can't quantify: "how safe is it" is a question with no numeric answer, so it loses to questions that do have one. Drishti's move is to translate the attack graph into a dollar figure — the expected cost of the most plausible path, weighted by likelihood. Suddenly the conversation shifts from "should we care about this CVE" to "this path costs more than the patch does, let's close it."
That's not a minor UX detail. It's the difference between a finding that lives in a ticket queue for a year and one that gets fixed this sprint.
Where LLMs genuinely help (and where they don't)
The honest take: the LLM is *not* the scanner. It doesn't discover vulnerabilities. What it's good at is the last, hardest mile — turning a correct technical conclusion into language a budget-holder acts on. The failure mode to avoid is letting it invent findings; the model drafts *remediation* over a *deterministic* attack graph, never the graph itself. The scanner is the source of truth; the LLM is the translator.
Takeaway
If you're building anything security-facing, build the translation layer. The scanning is a solved problem; the reason findings don't get fixed is a communication problem, and that's where LLMs and a priced, visual attack model actually move the needle.