Claude-OSINT: Turning an AI Into a Recon Operator (For Targets You’re Allowed to Touch)
Reconnaissance is where every engagement starts — before you touch anything, you map what’s there. Claude-OSINT is a set of skill files that turn Claude into a structured recon operator for that first phase. I rate it — with one condition I’m going to put before anything else, because it isn’t optional.
The condition: authorization. This is for assets you own or have written permission to assess — a bug-bounty scope, a signed red-team engagement, your own infrastructure. The project draws that line itself: it stops at OSINT-driven recon and deliberately excludes exploitation, post-exploitation and malware. Point it at anyone else and the problem is you, not the tool. Recon on a target you have no right to assess is still trespass.
What it is
Built by ElementalSoul (GenAI Security Research), released under the MIT licence. It’s about 4,600 lines of methodology and tradecraft split into two skills you drop into Claude:
- osint-methodology (~455 lines) — teaches Claude how to think: asset-graph discipline, severity rubrics, time-budgeting, and clean deliverable templates.
- offensive-osint (~4,168 lines) — teaches Claude what to reach for: probe wordlists, vendor fingerprints, dorks and reference patterns.
By the numbers, across 12 recon domains:
| Count | What | Notes |
|---|---|---|
| 90+ | recon modules | Identity/SSO, web surfaces, cloud, secrets, breach intel, supply chain and more |
| 80+ | code-search dorks | For finding exposed things in public code |
| 48 | secret-regex patterns | Spotting leaked keys — including modern AI and registry tokens |
| 9 | read-only credential validators | Check if a found key is live without using it destructively |
| 27 | attack-path templates | Structured escalation guidance for the report, not for firing |
The part I actually rate: it teaches judgment
Most “AI for hacking” projects are just a pile of tools. The failure mode of an LLM doing recon is confident noise — a hundred half-relevant findings with no sense of what matters. The methodology skill is the antidote: it makes the model build an asset graph, rank findings by severity, budget its time, and write up results in a consistent shape. That is the difference between a recon operator and a search-engine-with-vibes.
Read-only by design
The nine credential validators are the tell. They confirm whether a leaked key is live without doing anything destructive with it — the responsible version of “is this real?” And the whole kit stops at reconnaissance: no exploitation modules, no payloads that pop a box. For a tool this capable, that restraint is the right call, and it’s why I’m comfortable pointing you at it.
Where it fits your workflow
This is the pre-engagement, external, read-only phase — the Reconnaissance stage of the kill chain. It pairs naturally with the manual toolkit in our OSINT & recon cheat sheet, and with nmap once you move from “what exists” to “what’s open” on scope you’re cleared to scan.
The honest caveat
It is still an LLM. It will occasionally invent a subdomain, a service, or a finding that looks completely plausible and isn’t there. Every result needs verifying against the real source before it goes in a report — the model’s confidence is not evidence. This is the misinformation risk from the OWASP LLM Top 10, live and in your recon notes. Used well, Claude-OSINT makes you faster; it doesn’t make you right. You are accountable for the findings, not the model.
The project: elementalsouls/Claude-OSINT on GitHub (MIT, authorization-gated by design). Steven found it; I checked it and wrote this. Use it only on targets you’re allowed to touch. Something off in this write-up? Tell me — that link catches AI slop, dead links and bad takes alike.