ssrf
Skill2 starsChanged 4 months agoMITView on GitHub
SAST detection methodology for Server-Side Request Forgery (CWE-918), including blind/OOB SSRF, DNS rebinding, cloud-metadata credential theft (169.254.169.254), redirect-following and protocol-smuggling bypasses (gopher/file/dict), URL-parser confusion, and encoding/allowlist bypasses. Use when reviewing code that fetches a URL or opens a connection whose target is influenced by input. Writes confirmed findings to findings/24-ssrf.md.
12 other projects use a version of this.
Adopted by
Most recently taken up by
- mahumadad/quarrytheir own version, first seen Nov 8, 2026
- SubhiSec/claude-ai-cyber-security-skillstheir own version, first seen Oct 25, 2026
- ferr079/claude-code-cybersec-skillstheir own version, first seen Oct 12, 2026
- TaiMaBenJi/taima-skillstheir own version, first seen Oct 11, 2026
- cdxiaodong/cain-agenttheir own version, first seen Oct 8, 2026
- emre-guler/websectheir own version, first seen Aug 27, 2026
Used with
Projects with it also use sqli, ssti and 8 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
SubhiSecSkillsqli
Hunt SQL injection in REST/GraphQL/SOAP/raw parameters using sqlmap, ghauri, manual UNION/Time/Boolean techniques, and custom WAF-bypass tamper scripts. Use when an input flows into a database query or when the user has identified a parameter that might be vulnerable.
Also in 6 projects that use this
SubhiSecSkillssti
Hunt Server-Side Template Injection (SSTI) in Jinja2, Twig, Freemarker, Velocity, Smarty, ERB, Handlebars, Mustache, and Spring (SpEL). Use when input is reflected in a way that suggests template rendering — error pages, email templates, PDF generators, dynamic content systems.
Also in 7 projects that use this
zebbernSkillsubdomain-enum
Enumerate subdomains for a target domain using passive and active reconnaissance techniques. Use when starting reconnaissance on a new target, when you need to discover the attack surface, when expanding scope from a root domain, or when the user provides a domain to investigate.
Also in 4 projects that use this
langbyyiSkilldeserialization
Insecure deserialization playbook. Use when Java, PHP, or Python applications deserialize untrusted data via ObjectInputStream, unserialize, pickle, or similar mechanisms that may lead to RCE, file access, or privilege escalation.
Also in 4 projects that use this
langbyyiSkillpath-traversal
Path traversal and LFI playbook. Use when file paths, download endpoints, include operations, archive extraction, or wrapper behavior may expose filesystem control.
Also in 4 projects that use this
iAskShahramSkillfile-upload
Uploads a local file to Cloudflare R2 bucket `public` with wrangler and returns the public URL on assets.shahram.dev. Use when the user wants a public link, public URL, or to host/upload a local image or file to R2.
Also in 7 projects that use this
- +3 more from SubhiSec
Other versions12 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 12
- cdxiaodong/cain-agent★ 953same name
- SubhiSec/claude-ai-cyber-security-skills★ 45same name
- 0xGhostCAT/claude-ai-cyber-security-skills★ 41same name
- thejefflarson/soundcheck★ 20same name
- sunilgentyala/OmniRed★ 5same name
- langbyyi/websec-skills★ 4same name
- emre-guler/websec★ 2same name
- mahumadad/quarry★ 1same name
- TaiMaBenJi/taima-skills★ 1same name
- zyphr-labs/bounty-harness★ 1same name
2 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository49 other files it gives its agents.
- authentication-flawsSkill
- broken-access-control-idorSkill
- business-logic-flawsSkill
- cicd-and-pipeline-injectionSkill
- code-injectionSkill
- command-injectionSkill
- cors-misconfigurationSkill
- crlf-and-header-injectionSkill
- cross-site-scriptingSkill
- csrfSkill
- dependency-and-supply-chainSkill
- dom-clobbering-and-postmessageSkill
- graphql-injection-and-abuseSkill
- hardcoded-secretsSkill
- http-request-smugglingSkill
- iac-misconfigurationSkill
- information-disclosureSkill
- insecure-deserializationSkill
- insecure-randomnessSkill
- jwt-vulnerabilitiesSkill
- ldap-injectionSkill
- llm-and-ai-application-securitySkill
- mass-assignmentSkill
- memory-safetySkill
- mobile-android-securitySkill
- mobile-ios-securitySkill
- nosql-injectionSkill
- oauth-oidc-saml-flawsSkill
- open-redirectSkill
- path-traversal-and-file-inclusionSkill
- prototype-pollutionSkill
- race-conditions-and-toctouSkill
- redosSkill
- sast-final-reportSkill
- sast-finding-validationSkill
- sast-methodologySkill
- security-logging-and-monitoring-failuresSkill
- security-misconfiguration-and-headersSkill
- server-side-template-injectionSkill
- session-managementSkill
- smart-contract-securitySkill
- sql-injectionSkill
- unrestricted-file-uploadSkill
- unsafe-parsers-and-yamlSkill
- weak-cryptographySkill
- web-cache-poisoning-and-deceptionSkill
- websocket-securitySkill
- xpath-and-xml-injectionSkill
- xxeSkill
The file
We split this skill into the rules it gives an agent
5 rules. 0 are common, 3 are distinctive, and 2 are very specific to this project.
0
Common
Most of the 10 edited versions we read keep these.
3
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 34
Document/media renderers that dereference URLs: PDF/HTML-to-PDF (wkhtmltopdf, headless Chrome/Puppeteer), image proxies and thumbnailers, …
- Very specific: line 43
All HTTP inputs, but especially any field that *is* or *becomes* a URL/host/port: ?url=, JSON body URLs, webhook/callback config, …
- Distinctive: line 101
Prefer a mapping table (user picks an ID → server holds the real URL) over fetching a raw user URL at all.
- Very specific: line 194
Medium: when egress is provably confined to an isolated network with no sensitive internal targets.
- Distinctive: line 199
Do not fetch raw user URLs — map an ID to a server-held URL where possible.
“Common” is measured: kept by at least 6 in 10 of the 10 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Server-Side Request Forgery — SAST Methodology
Class: SSRF · CWE-918 · OWASP: A10 Server-Side Request Forgery
Findings file: findings/24-ssrf.md
1. Overview
SSRF occurs when attacker-controlled data determines the destination of a request the server makes — the host, port, scheme, or any component that resolves to a network endpoint. The server becomes a proxy into places the attacker cannot reach directly: the loopback interface, internal services, the cloud metadata endpoint, or arbitrary external hosts. The core test: does a source value control (any part of) the target of an outbound request, and is the final resolved destination constrained by an allowlist enforced at connection time? Anything weaker — string checks, blocklists, validating a URL that is then redirected or re-resolved — is bypassable.
2. Where it lives
- Any "fetch this URL for me" feature: webhooks, import-from-URL, avatar/logo-from- URL, RSS/feed readers, link previews / URL unfurling, OAuth/OIDC discovery and JWKS fetch, SSO/SAML metadata loaders, remote-file mirrors.
- Document/media renderers that dereference URLs: PDF/HTML-to-PDF (wkhtmltopdf, headless Chrome/Puppeteer), image proxies and thumbnailers, SVG rasterizers, XSLT/XML processors (overlaps XXE), office-doc converters.
- Server-side HTTP clients driven by config the user can set (webhook target, callback URL, proxy host, S3-compatible endpoint).
- "Health check", "test connection", and open-proxy/gateway endpoints.
3. Sources (tainted input)
All HTTP inputs, but especially any field that is or becomes a URL/host/port:
?url=, JSON body URLs, webhook/callback config, Location-style redirect
targets, a hostname assembled from parts. Second-order is common and
high-miss: a webhook URL or avatar URL stored at configuration time and fetched by
a background job later — trace the stored value back to its write. Also indirect:
a filename or ID that is concatenated into an internal URL, or a Host/X-
Forwarded-Host header reflected into a server-side fetch.
4. Sinks (dangerous operations)
# Python — dangerous
requests.get(user_url) # host fully attacker-controlled
urllib.request.urlopen(url) # also honors file://, ftp://
httpx.get(url, follow_redirects=True) # redirect can jump to 169.254.169.254
requests.get(url, allow_redirects=True) # validate-then-redirect bypass// Node — dangerous
await fetch(req.query.url); // no destination allowlist
axios.get(userUrl); // follows redirects by default
http.get(new URL(target)); // SSRF into internal services
require('request')({ url: target, followRedirect: true });// Java — dangerous
new URL(userUrl).openConnection().getInputStream();
restTemplate.getForObject(userUrl, String.class); // follows redirects
HttpClient.newHttpClient().send(HttpRequest.newBuilder(URI.create(u)).build(), ...);// Go — dangerous
http.Get(userURL) // default client follows redirects// PHP — dangerous
file_get_contents($_GET['url']); // wrappers: php://, file://, gopher://
curl_setopt($ch, CURLOPT_URL, $_GET['url']); // + CURLOPT_FOLLOWLOCATIONDangerous whenever the host/URL is source-derived and the client will connect to
whatever it resolves to — including after a redirect, and including non-http
schemes (file://, gopher://, dict://, ftp://) that some clients accept.
5. Sanitizers / safe patterns
Safe:
- Allowlist the destination host, and enforce it against the resolved IP at
connection time, not the URL string. Resolve the hostname once, reject the IP
if it is private/loopback/link-local/reserved (
127.0.0.0/8,10/8,172.16/12,192.168/16,169.254/16,::1,fc00::/7,0.0.0.0), then connect to that pinned IP so the name cannot re-resolve (defeats DNS rebinding). - Restrict the scheme to
http/httpsonly; forbid redirects, or re-validate the target on every hop. - Route outbound fetches through a dedicated egress proxy / isolated network with no route to metadata or internal ranges; disable the cloud metadata endpoint or require IMDSv2 (session-token) so a bare GET fails.
- Prefer a mapping table (user picks an ID → server holds the real URL) over fetching a raw user URL at all.
Fails / not a real sanitizer:
- Blocklisting hostnames/IPs by string — bypassed by
127.0.0.1alternates: decimal2130706433, octal0177.0.0.1, hex0x7f000001,0,[::],127.1, enclosed-alphanumeric/Unicode digits, andlocaltest.me-style names that resolve to loopback. - Validate-then-fetch with redirects on — the URL passes the check, then a
302sends the client tohttp://169.254.169.254/. TOCTOU. - DNS rebinding — the name resolves to a public IP during validation, then to
169.254.169.254/internal at connection time (low TTL). Any design that resolves twice is vulnerable. - Parser confusion — the validator and the HTTP client disagree on the host.
http://expected.com@169.254.169.254/,http://169.254.169.254#@expected.com/,http://expected.com\@evil, backslash/whitespace tricks; one library reads the host asexpected.com, the other connects to the attacker's host. - Allowlist by substring (
if "expected.com" in url) —expected.com.evil.tldorevil.tld/expected.compasses.
6. Detection methodology
- Find sinks: grep for outbound HTTP/URL-open APIs.
rg -n 'requests\.(get|post|put|head|request)\(|urlopen\(|httpx\.|urllib' rg -n 'fetch\(|axios\.|http\.get\(|https\.get\(|got\(|\brequest\(' rg -n 'new URL\(|openConnection|HttpClient|RestTemplate|WebClient|http\.Get\(' rg -n 'file_get_contents\(|curl_setopt|CURLOPT_URL|fopen\(|fsockopen' - Confirm the target is source-derived: is the URL/host argument a variable traceable to request input or stored config, not a constant?
- Trace redirect and re-resolution behavior: does the client follow redirects? Is the host resolved separately from where the connection actually goes?
- Check for a real destination control: allowlist enforced on the resolved IP at connect time? Or only string checks / blocklists (see §5)?
- Confirm reachability: exposed route/webhook/job; pre-auth? Note whether the response is returned to the attacker (in-band) or not (blind).
- Classify: in-band vs blind/OOB; reachable targets (loopback, internal range, metadata); usable non-http schemes.
7. Modern & niche variants
- Blind / OOB SSRF: the response is never returned. Still fully exploitable — confirm via out-of-band interaction (attacker-controlled DNS/HTTP callback) and exploit timing/side effects. Common in webhook, PDF-render, and background-fetch paths. Do not dismiss because "the output isn't shown."
- DNS rebinding (TOCTOU on resolution): validation resolves the hostname to a
benign IP; the HTTP client re-resolves (low TTL) to
169.254.169.254or an internal IP at connection time. The fix is to resolve once and connect to the pinned IP — flag any validate-then-connect design that resolves twice. - Cloud metadata → credentials (the key chain):
http://169.254.169.254/…on AWS IMDSv1 returns temporary IAM credentials (/latest/meta-data/iam/security-credentials/<role>); GCP requires headerMetadata-Flavor: Googleatmetadata.google.internal; Azure uses169.254.169.254/metadata/instance?api-version=…withMetadata: true. SSRF that reaches these turns into stolen cloud credentials → full account pivot. This is why metadata reachability elevates severity to Critical. - Redirect-following bypass: the URL is validated but the client follows a
30xto an internal/metadata host. Disable redirects or re-validate each hop. - Protocol smuggling: clients that honor
gopher://,file://,dict://,ftp://let the attacker read local files (file:///etc/passwd) or craft raw TCP payloads.gopher://127.0.0.1:6379/_…speaks the Redis line protocol (write keys, schedule cron → RCE);dict:///gopher://also hit unauthenticated internal services (Memcached, Elasticsearch, internal HTTP). - URL-parser confusion: validator and fetcher disagree on the host —
user@host,#/?fragment tricks, backslashes, embedded whitespace/tabs, and differing WHATWG-vs-RFC parsing. A prime allowlist bypass. - Encoding bypasses for the host: decimal/octal/hex IP forms,
0x7f.1, IPv6[::ffff:127.0.0.1]and[::1], IPv4-mapped IPv6, enclosed-alphanumeric and Unicode digit look-alikes, and short forms (127.1,0). - Allowlist bypass via
@/#/?:https://allowed.com@attacker/,https://attacker/#allowed.com,https://attacker/?x=allowed.com— the attacker's host is the real authority while the check seesallowed.com. - Feature-driven SSRF: webhooks, PDF/HTML renderers, image proxies, import-
from-URL, SVG (external
<image href>/xlink), and link previews are the most common real-world entry points — each fetches a URL by design.
8. Common false positives
- The target host is a hard-coded constant or selected from a fixed server-side map (user only picks an ID).
- Fetch is confined to a dedicated egress proxy with no route to internal ranges or metadata, verified in config.
- Destination is allowlisted and enforced on the resolved IP at connect time with redirects disabled.
- Client library restricted to
httpsto a single external SaaS with pinned host.
9. Severity & exploitability
Base High (reaching internal services). Critical when it reaches cloud
metadata and yields credentials (metadata → creds → account pivot), enables
protocol smuggling to an unauthenticated internal service (Redis/gopher → RCE), or
is pre-auth. Blind SSRF keeps High if internal reachability is proven. Lower to
Medium when egress is provably confined to an isolated network with no
sensitive internal targets. See references/severity-model.md.
10. Remediation
Do not fetch raw user URLs — map an ID to a server-held URL where possible. Where a
URL must be fetched: enforce a strict destination allowlist against the resolved
IP at connection time, pin that IP for the actual connection (defeat rebinding),
restrict schemes to http/https, and disable or re-validate redirects. Isolate
outbound fetches on a network with no route to internal ranges or metadata, and
require IMDSv2. Never rely on hostname blocklists or string checks.
11. Output
Append each confirmed finding to findings/24-ssrf.md using
references/finding-template.md. Set Class: SSRF, CWE: CWE-918, and in
Chain potential name primitives such as internal service access (→ pivot to
unauthenticated internal APIs), cloud credential theft (metadata → IAM creds,
provides leaked-credentials for privilege-escalation/lateral chains), arbitrary
file read (file://), and internal RCE (gopher:// → Redis). Note whether the
finding consumes an open-redirect or allowlist-bypass primitive from another
finding.
Primitives (controlled): provides OUTBOUND_REQUEST,SECRET_LEAK (metadata),FILE_READ (file://); consumes URL_CONTROL (redirect/allowlist bypass)
References
- CWE-918 Server-Side Request Forgery; OWASP A10:2021 SSRF; OWASP SSRF Prevention Cheat Sheet; cloud provider IMDS hardening guides (AWS IMDSv2, GCP/Azure metadata headers).
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

