Security agents inspect a codebase and save what they learn as taint rules and dependency models. On later scans, static analysis applies those files instead of asking an LLM to rediscover the same behavior. This makes agentic application security testing continuous and lean.
OpenTaint is the static-analysis part of this workflow. It is an open source taint analysis engine that tracks untrusted data across methods, fields, aliases, asynchronous code, persistence layers, and framework boundaries.
During a security review, the agent creates two kinds of files:
- Taint rules define vulnerability classes using sources, sinks, and sanitizers.
- Dependency models describe how data moves through library and framework code that OpenTaint does not analyze directly.
OpenTaint applies both across the whole codebase. When a scan misses a flow, the agent adds or corrects a rule or dependency model. When a scan reports a safe flow, the agent adds a sanitizer. These changes remain in the project, so every later scan uses them. The agent learns a behavior once, and OpenTaint checks it continuously.
The engine, CLI, rules, and CI integrations are available under Apache 2.0 and MIT licenses.
Run a scan without installing anything:
npx @seqra/opentaint scanRead the quick-start guide, or install the agent skills for an end-to-end application-security review.
See the agent workflow
The demo below shows an agent running OpenTaint, following the reported flows, and triaging the results.
How OpenTaint balances accuracy, precision, and performance
Every static-analysis engine must balance three goals:
- Accuracy — find real vulnerabilities and avoid false negatives.
- Precision — report real issues and avoid false positives.
- Performance — finish with reasonable time and memory on large codebases.
Improving one goal by simplifying the analysis often weakens another. For example, merging all fields of an object can make analysis faster, but it may report safe fields or lose a dangerous flow.
OpenTaint propagates taint to a fixpoint instead of stopping at a fixed depth. It tracks fields separately, resolves aliases and pointers, and keeps calling contexts separate. Extended IFDS, abductive inference, and compact taint representations control the time and memory required for that analysis.
Supported languages and roadmap
OpenTaint supports Java and Kotlin today, with deep modeling for Spring applications.
Next: Python and Go.
Planned: C#, JavaScript, and TypeScript.
To report false positives or missed flows, improve rules, submit dependency models, or help extend the engine, open an issue or join us on Discord.
