Taint analysis

Sources, sinks, sanitizers, guards and cross-file traces.

Sources

A source is where untrusted data enters: request parameters, headers, cookies and bodies in each framework; command-line arguments, environment variables, stdin and sockets in C/C++; Android intent extras and deep-link URIs; values returned by helper methods (in the same file or, for Java, in another file of the scan) that themselves return request data.

Sinks

A sink is a call where attacker-controlled data causes harm: SQL execution, shell commands, file paths, outbound URLs, HTML output, deserializers, template compilers, LDAP and XPath queries, raw memory operations and so on. Each sink is matched on the specific argument that matters — the SQL text, not the bound parameters; the program name and a shell -c string, not other arguments.

Sink classes and sanitizers

A tracked value carries a set of vulnerability classes rather than a single "tainted" flag. A sanitizer clears only the classes it actually neutralises: HTML escaping clears XSS but not SQL injection, so an HTML-escaped value reaching a SQL query is still reported, and the trace says so. Converting to a number clears the injection classes but keeps the fact that an attacker chose the value — which is what matters for an unbounded allocation or an out-of-range index.

Guards

Only unambiguous checks clear a value: membership in a literal allow-list, a fully anchored character-class pattern, equality with a literal, a bound check against a constant or a length, a resolved path checked against its base directory. A loose regular-expression match is deliberately not trusted.

Paths and functions

Branches are analysed separately and merged (a value sanitised on one branch and raw on the other stays reported). Function summaries record which parameters reach a sink or the return value, and are iterated to a fixed point so long helper chains resolve.

Across files

Imports, injected services, packages and includes are followed in all ten languages. A flow that crosses files is reported once, at the caller, with every file on its path in the trace. For C#, a public helper that other files only ever call with constants is proven safe from those call sites.

Reading a trace

Each step has a kind — source, assignment, call, sanitizer, cross-file, parameter, sink — a file and line, a short label and the code snippet. The live demo shows real traces.

What sets the engines apart

  • Source → Sink Traces. Every finding ships with the actual data-flow path, hop by hop, with real file and line numbers — including across a file boundary.
  • Cross-File Data Flow. In all ten languages: imports and re-exports, injected services (Spring, ASP.NET, NestJS, Axum state), Go packages, PHP includes, Rails autoloading, Rust modules — with a 72-case cross-file suite across six of the languages run on every change.
  • One Issue, One Finding. When the pattern layer, the taint engine and cross-file analysis see the same bug, it becomes one finding listing every location and every detector that agreed.
  • Broken Object-Level Authorization. Structural ownership-dominance analysis: a resource lookup only clears when a real comparison against the principal dominates every path to the sink.
  • Reachable Dependencies. Live CVEs across 8 ecosystems, ranked by whether the vulnerable function is called, the package is reachable from an entry point, used only in tests, or never imported.
  • Cloud & IaC Posture. Terraform (AWS, GCP, Azure), CloudFormation, ARM, Bicep and Serverless: public storage, open admin ports, public databases, wildcard IAM, hard-coded credentials — no cloud credentials needed.
  • Kubernetes & Containers. Per-container hardening in manifests and Helm charts, RBAC wildcards, committed Secrets; end-of-life base images and disabled sandboxing in Dockerfiles and Compose.
  • API Endpoint Inventory. Every route across Express, Next.js, Flask, FastAPI, Spring, ASP.NET, Go and Laravel plus OpenAPI specs — and the endpoint missing the auth its siblings have.
  • Stable Finding Fingerprints. A hash of the flow itself, not the line number — the same finding survives unrelated edits elsewhere in the file across scans.
  • Reviewer Attestation. Named sign-off recorded per file with signature, timestamp, and risk context at review time — an audit trail that answers itself.
  • Context-Aware Sanitizers. An HTML-escaped value dropped into a <script> block or an unquoted attribute is still flagged — encoding for one context doesn't cover another.
  • Exact Sink Arguments. Only the exploitable argument counts: the SQL text, not bound parameters; the LDAP filter, not the base DN; the program, not its arguments.

Next: AI attribution →