Irify Sast

IRify SAST

Deep static analysis skill powered by IRify's SSA compiler and SyntaxFlow query engine.

Prerequisites

This skill requires the yaklang MCP server. Configure it in your agent's MCP settings:

# Codex: ~/.codex/config.toml
[mcp_servers.yaklang-ssa]
command = "yak"
args = ["mcp", "-t", "ssa"]

// Claude Code / Cursor / others
{ "command": "yak", "args": ["mcp", "-t", "ssa"] }

Workflow: Engine-First (sf → read → rg)

CRITICAL: Always follow the Engine-First funnel model. The SSA engine sees cross-procedure data flow across all files simultaneously — grep cannot. Do NOT use grep/rg to build a "candidate file pool" before querying. Instead, let the engine be your radar first.

Step 1: Compile (once per project, auto-cached)

ssa_compile(target="/path/to/project", language="java", program_name="MyProject")
→ full compilation, returns program_name

Auto Cache: If the program was already compiled and source files haven't changed, the engine returns [Cache Hit] instantly — no recompilation. Always provide a program_name to enable caching.

Step 2: Query — use SyntaxFlow as the global radar

Directly compose and execute SyntaxFlow rules against the compiled IR. Do NOT pre-scan with grep to find candidates.

ssa_query(program_name="MyProject", rule="<SyntaxFlow rule>")

The engine traverses the entire SSA graph in memory, crossing all file boundaries. One query covers what would take dozens of grep commands, with zero false positives on data flow.

Step 3: Read — use Read as the microscope

After ssa_query returns concrete file paths and line numbers, use Read to examine surrounding context (±20 lines). Verify whether the hit is real business code or dead/test code.

Step 4: Grep — use Grep/Glob only for non-code files

Use Grep/Glob only for content that the SSA engine does not process:

• Configuration files (.yml, .xml, .properties, logback.xml)
• Static resources, templates, build scripts
• Quick name/path lookups when you already know the exact string

NEVER use Grep to search for data flow patterns in source code — that is what ssa_query is for.

Incremental Compile (when code changes)

ssa_compile(target="/path/to/project", language="java", base_program_name="MyProject")
→ only changed files recompiled, ProgramOverLay merges base + diff layers
→ returns NEW program_name for subsequent queries

IMPORTANT: Use base_program_name for incremental compilation. re_compile=true is a full recompile that discards all data — only use it to start completely fresh.

Self-Healing Query (auto-retry on syntax error)

When ssa_query returns a SyntaxFlow parsing error:

1. DO NOT apologize to the user or ask for help
2. Read the error message — it contains the exact parse error position and expected tokens
3. Fix the SyntaxFlow rule based on the error
4. Re-invoke ssa_query with the corrected rule
5. Repeat up to 3 times before reporting failure
6. If all retries fail, show the user: the original rule, each attempted fix, and the final error

Critical: Follow User Intent

DO NOT automatically construct source→sink vulnerability rules unless the user explicitly asks for vulnerability detection.

• User asks "find user inputs" → write a source-only rule, list all input endpoints
• User asks "find SQL injection" → write a source→sink taint rule
• User asks "where does this value go" → write a forward trace (-->) rule
• User asks "what calls this function" → write a call-site rule

Source-Only Query Examples (Java)

When the user asks about user inputs, HTTP endpoints, or controllable parameters:

// Find all Spring MVC controller handler methods
*Mapping.__ref__?{opcode: function} as $endpoints;
alert $endpoints;

// Find all user-controllable parameters in Spring controllers
*Mapping.__ref__?{opcode: function}<getFormalParams>?{opcode: param && !have: this} as $params;
alert $params;

// Find GetMapping vs PostMapping endpoints separately
GetMapping.__ref__?{opcode: function} as $getEndpoints;
PostMapping.__ref__?{opcode: function} as $postEndpoints;
alert $getEndpoints;
alert $postEndpoints;

Source→Sink Query Examples (only when user asks for vulnerability detection)

// RCE: trace user input to exec()
Runtime.getRuntime().exec(* #-> * as $source) as $sink;
alert $sink for {title: "RCE", level: "high"};

// SQL Injection (MyBatis): detect ${} unsafe interpolation in XML mappers / annotations
// <mybatisSink> is a dedicated NativeCall that finds all MyBatis ${} injection points
<mybatisSink> as $sink;
$sink#{
    until: `* & $source`,
}-> as $result;
alert $result for {title: "SQLi-MyBatis", level: "high"};

Proactive Security Insights

After running a query and finding results, proactively raise follow-up questions and suggestions. Do NOT just dump results and stop.

When vulnerabilities are found:

1. Suggest fix: "This exec() call receives unsanitized user input. Consider using a whitelist or ProcessBuilder with explicit argument separation."
2. Ask related questions:
   • "Should I check if there are other endpoints that also call Runtime.exec()?"
   • "Want me to trace whether any input validation/sanitization exists between the source and sink?"
   • "Should I look for similar patterns in other controllers?"
3. Cross-reference: If one vulnerability type is found, proactively scan for related types:
   • Found RCE → "I also checked for SSRF and found 2 potential issues. Want details?"

When no results are found:

1. Don't just say "no results" — explain WHY:
   • "No direct exec() calls found, but I see ProcessBuilder usage. Want me to check those instead?"
   • "The query matched 0 sinks. This could mean the code uses a framework abstraction — want me to search for framework-specific patterns?"
2. Suggest alternative queries

When results are ambiguous:

1. Ask for clarification: "I found 8 data flow paths to executeQuery(), but 5 use parameterized queries (safe). Want me to filter to only the 3 using string concatenation?"

Companion Reference Files

When writing SyntaxFlow rules, read these files using the Read tool for syntax help and real-world examples:

Workflow:

1. Read syntaxflow-examples.md to find a similar rule pattern
2. Need a NativeCall? Read nativecall-reference.md
3. Compose and execute via ssa_query

SyntaxFlow Quick Reference

Search & Match

documentBuilder          // variable name
.parse                   // method name (dot prefix)
documentBuilder.parse    // chain
*config*                 // glob pattern
/(get[A-Z].*)/           // regex pattern

Function Call & Parameters

.exec()                           // match any call
.exec(* as $params)               // capture all params
.parse(*<slice(index=1)> as $a1)  // capture by index

Data Flow Operators

.exec(* #-> * as $source)            // trace param origin
$userInput --> as $sinks              // trace where value goes
$sink #{depth: 5}-> as $source       // depth-limited trace
$val #{
  include: `*?{opcode: const}`
}-> as $constSources                  // filter during trace
$sink #{
  until: `* & $source`,              // stop when reaching source
}-> as $reachable

Filters ?{...}

$vals?{opcode: call}                // by opcode: call/const/param/phi/function/return
$vals?{have: 'password'}            // by string content
$vals?{!opcode: const}              // negation
$vals?{opcode: call && have: 'sql'} // combined
$factory?{!(.setFeature)}           // method NOT called on value

Variable, Check & Alert

.exec() as $sink;                                      // assign
check $sink then "found" else "not found";             // assert
alert $sink for { title: "RCE", level: "high" };       // mark finding
$a + $b as $merged;                                    // union
$all - $safe as $vuln;                                 // difference

NativeCall (40+ built-in functions)

Most commonly used — see nativecall-reference.md for full list:

<include('rule-name')>         // import lib rule
<typeName()>                   // get short type name
<fullTypeName()>               // get full qualified type name
<getReturns>                   // function return values
<getFormalParams>              // function parameters
<getFunc>                      // enclosing function
<getCall>                      // find call sites
<getCallee>                    // get called function
<getObject>                    // parent object
<getMembers>                   // object members
<name>                         // get name
<slice(index=N)>               // extract by index
<mybatisSink>                  // MyBatis SQL injection sinks
<dataflow(include=`...`)>      // filter data flow paths

Tips

1. #-> = "where does this come from?", --> = "where does this go?"
2. Use * for params, don't hardcode names
3. SSA resolves assignments: a = getRuntime(); a.exec(cmd) = getRuntime().exec(cmd)
4. Use opcode filters to distinguish constants / parameters / calls
5. Combine check + alert for actionable results
6. After code changes, use base_program_name (not re_compile) for fast incremental updates
7. Before writing a new rule, read syntaxflow-examples.md to find similar patterns
8. When unsure about a NativeCall, read nativecall-reference.md for usage and examples