skillshop-ostyles/agents-security

API contract guard: extracts the API surface (HTTP routes with params, DTO fields, exported signatures - preferring OpenAPI files when present) from two git states of a repo, diffs them, classifies every change as breaking / non-breaking / additive, and writes a ready-to-ship consumer migration note per breaking change. Read-only. Trigger: /api-diff

Authorization X-ray for your own codebase (defensive audit): inventories every HTTP endpoint and every recognizable protection layer (middleware chains, authorize decorators, inline role checks, router mounts), builds the permission matrix endpoint x required check, and reports unprotected mutating endpoints and inconsistent protection of similar resources. Static, sends no requests. Read-only. Trigger: /authz

Finds mutating endpoints that lack explicit authorization, relying solely on middleware inheritance - the dangerous gaps where middleware failure leaves endpoints unprotected. Read-only. Audience: Senior. Trigger: /authz-coverage

Configuration cartographer: maps a system's complete config surface - every env var, setting and flag, where it is defined (.env, yaml/json configs, compose, Dockerfile) versus where it is read in code - and reports read-but-never-defined keys (crash candidates), defined-but-never-read orphans and divergent defaults. Never outputs values, keys only. Read-only. Trigger: /config-map

CORS config drift scanner: harvests every Access-Control-Allow-Origin header, cors()-middleware call, @cross_origin decorator, options-handler with cors config, and per-route origin/credentials settings. LLM analyses each per-route CORS posture: credentials+wildcard = fatal, permissive origin patterns, preflight gaps. Read-only. Audience: Senior. Trigger: /cors-drift

Crypto downgrade detector: harvests every weak-algorithm usage (MD5, SHA1, DES, 3DES, RC4, ECB, CBC, deprecated createCipher), modern alternatives (subtle.encrypt, bcrypt, argon2, scrypt, PBKDF2), JWT signing-config (hardcoded-secret vs asymmetric key), cert/key-generation calls, and patch/version-based crypto weakeners (ALLOW_WEAK, --harmony, legacy-crypto). LLM analyses each finding as downgradeable, deprecated, or acceptable and recommends minimum upgrade. Read-only. Audience: Senior. Trigger: /crypto-downgrade

Maps PII fields and their sinks - logs, third-party APIs, exports - purely via field names, never via actual data. Trigger: /data-trail-tracker

Dependency inheritance audit: for every direct dependency answers the questions nobody asks - why is it here (from actual usage sites), how deep is the coupling, how replaceable is it, and what is the concrete exit plan. Parses manifests/lockfiles, scans usage, optionally enriches with registry metadata (offline-safe). Read-only. Trigger: /deps-audit

Error message leakage detector: harvests every HTTP-error-return and log-error-call, classifies what kind of information leaks (stacktrace, SQL error message, env-vars, user input echo, request dump). LLM validates each finding as legitimate production-leak and proposes sanitization. Read-only. Audience: Both. Trigger: /error-leakage

Flask anti-pattern detector: scans Flask projects for hardcoded SECRET_KEY, debug-mode in production, dangerous template rendering (render_template_string), pickle/eval/exec on request data, unsafe session config, SQL injection via raw queries, insecure file upload, and debug toolbar enabled. LLM validates each finding and proposes modern alternatives. Read-only. Audience: Senior. Trigger: /flask-detector

Input validation audit: statically detects all input surfaces (HTTP params, CLI args, env vars, file reads, stdin) across a codebase, classifies their validation state (none/weak/adequate), and flags high-risk gaps. Produces an evidence-backed report with severity, location, and remediation suggestions. Read-only. Trigger: /input-audit

Log injection detector: harvests every console.log/logger.info/log.Error/etc call, classifies arguments for attacker-controlled input (CWE-117), CRLF injection surface, sensitive data leakage (passwords, tokens, secrets). LLM validates each finding as injectable and proposes sanitization (parameterized logging, newline stripping, sensitive-field redaction). Read-only. Audience: Senior. Trigger: /log-injection

Permission chain analyzer: extracts role definitions, role check sites, middleware mounts, and mutating routes. Identifies transitive chains (role A can reach endpoint E via routes R1,R2...), surfaces unprotected mutating routes (file-local check missing), detects divergent role-naming (same role defined differently in 3 files). Read-only. Audience: Senior. Trigger: /permission-chain

Rate-limit shape analyzer: inventories rate-limit decorators per-route (express-rate-limit, flask-limiter, DRF throttle, Spring). Per endpoint, classifies whether it has a decorator, on what limit (max, window, per-tier), and whether mutating endpoints that should be limited are. Read-only. Audience: Senior. Trigger: /rate-shape

Secret lifecycle auditor: inventories every secret-shaped key/value across .env, k8s manifests, Vault configs, IAM refs, terraform. Per secret: age from git log -S, masked value (first-8/last-4), reachability-check against installed dependencies, type guess from key prefix. LLM judges rotation cadence and emits prioritized rotate-now / rotate-soon / remove-dead list. Read-only. Audience: Senior. Trigger: /secret-lifecycle

Security smell scanner: statically detects 10 families of security anti-patterns across a codebase (SQL injection, XSS, command injection, path traversal, hardcoded credentials, insecure defaults, IDOR, open redirect, TOCTOU, missing input validation). Produces an evidence-backed report with severity, location, and contextual analysis. Read-only. Audience: Senior > Vibe. Cross-link from quality/ cluster. Trigger: /security-scan

Session state anomaly scanner: finds every session-generation, session-id usage, post-auth session-regeneration, post-logout cleanup, and refresh-token rotation site. LLM per finding: regen? invalidate? rotated? What attack arises if not? Read-only. Audience: Senior. Trigger: /session-anomaly

SSRF detector: finds every outbound HTTP call (fetch, axios, got, http, requests, HttpClient, Invoke-RestMethod) where the URL is user-controlled (req.body/req.query/req.params) and grades URL-pre-fetch validation (URL-parse, hostname allowlist, metadata-IP blocking). LLM per-URL-flow classifies user-control, pre-validation quality, and metadata-service exploitation risk (169.254.169.254). Read-only. Audience: Senior. Trigger: /ssrf-detector

Third-party trust boundary analyzer: inventories every outbound HTTP/RPC call (fetch, axios, got, requests, curl, Invoke-RestMethod). For each, identifies literal-vs-template URL, classifies known-trusted vs unknown domain, detects auth-header presence in call window, flags webhook handlers missing signature verification. Read-only. Audience: Senior. Trigger: /third-party-trust

TLS config drift scanner: harvests every TLS-version constant, cipher-suite array, cert-pinning call, mTLS flag, cert-validation callback, and FIPS-mode setting. LLM analyses each statement as acceptable, misconfigured, or downgrade-prone when an element is missing. Read-only. Audience: Senior. Trigger: /ssl-drift

Type confusion bypass detector: traces validation paths from input source to storage/execution sink, tests edge-case input shapes (str/int/obj/null/array), and LLM judges which input shape circumvents each validator and what happens at the query/execution sink. Read-only. Audience: Senior. Trigger: /bypass-detector