elementalsouls/Claude-OSINT

What it is

claude-osint (by Sachin Sharma) is a library of 8 "skills" for Anthropic's Claude skills system, aimed at authorized red-team and bug-bounty external recon. Each skill is a structured SKILL.md file (~10,000 lines total) that primes Claude with methodology, wordlists, regexes, and validators for one part of offensive reconnaissance. It also ships two Python scripts and four end-to-end engagement walk-throughs.

How it works

Skills are dropped into ~/.claude/skills/, where Claude auto-triggers them on relevant phrases. The design is a two-part core plus six depth skills:

• osint-methodology (515 lines) — "how to think": asset-graph discipline, severity rubric, time budgeting, reporting templates, OpSec/detectability.
• offensive-osint (4,555 lines) — "what to reach for": probe wordlists, vendor fingerprints, cloud/K8s/CI-CD notes, a secret catalog, read-only validators, curl one-liners. Includes scripts/secret_scan.py (stdlib-only, 80 patterns) and scripts/h1_reference.py (HackerOne disclosed-reports reference).

The six organization-grade skills extend from single-target to enterprise attack-surface work: org-attack-surface (GLEIF org-tree, RIR netblock recall, ASN scope guard — discover-only), email-domain-security (spoofability verdict + SPF supply-chain), exposure-risk-quantification (FAIR 0–100 / A–F score, $-loss model, board one-pager), continuous-exposure-monitoring (re-scan/diff + CTI chatter), cloud-saas-exposure (bucket ownership-gated severity, offline AWS account-ID decode, dependency-confusion), and identity-provider-recon (tenant/federation mapping + pre-auth user-enumeration oracle).

Preconditions / Affected

Requires a Claude environment that supports the skills system and the ~/.claude/skills/ directory. Intended scope is explicitly authorized red-team and bug-bounty engagements.

What to do

If evaluating for use, note the built-in scope constraints, which are the security-relevant part: several skills bound themselves (org-attack-surface is "discover-only," identity-provider-recon states a "hard enumeration boundary," and validators are "read-only"). It also claims ownership/proof honesty caps and finding-lifecycle FP discipline. Treat the credential validators and user-enumeration oracle as the components most likely to cross a line if used outside an authorized engagement — confirm scope before triggering them.

Caveats

The metrics are self-reported: "56/56 PASS on a 56-prompt self-evaluation (100%, zero fabrications)" and "~85–90% practitioner coverage" both come from the project's own smoke tests (tests/smoke-test-prompts.md), not independent evaluation. "Zero fabrications" and the enumeration/discover-only boundaries are design claims, not verified guarantees — the actual behavior depends on how Claude interprets each SKILL.md at runtime.