Elo backends have code injection via unvalidated emission of programmatic AST fields

Summary

Elo's programmatic AST API accepts arbitrary values in many node fields without runtime type or syntax validation. Each backend's emit stage interpolates those values straight into the generated source with no escape pass. An integrator that populates the AST from JSON-deserialised input, plugin output, or any other untyped source therefore exposes a code-injection primitive that flows through to the emitted JavaScript, Ruby, Python, or SQL.

The same pattern was previously addressed in 0.9.7 (2026-01-27), which patched the date/datetime/duration literal emitters via JSON.stringify() after a report by naoyashiga. This advisory generalises that fix to the remaining AST surface — at least 23 emission sites across six node-field categories and four backends.

Vulnerable sinks

Each row is a category of sinks that share the same shape: TypeScript signature erased at runtime → AST-API consumer passes a crafted value → emit interpolates the value verbatim.

#	AST field	Affected backends	Resulting injection
1	int_literal / float_literal .value	JS, Ruby, Python, SQL	Arbitrary expression in the emitted numeric position (RCE on JS/Ruby/Python; SQL expression injection)
2	variable.name	JS, Ruby, Python, SQL	Identifier-position injection; on SQL, raw SQL injection because undefined identifiers are intentionally emitted as column references
3	member_access.property	JS, Ruby, Python, SQL	Property access escaped into expression (RCE)
4	object_literal key	JS, Ruby, SQL	Computed-key / extra-property injection (RCE on JS/Ruby; SQL injection)
5	datapath segment	JS, Ruby, Python, SQL	Ruby symbol-list injection allows backtick shell exec; SQL string-segment unquoted in single-quote context allows SQL injection
6	lambda.params[].name	JS, Ruby, Python	Parameter-list structural injection (RCE)
7	SQL let binding name	SQL	Identifier injection in … AS <name> clause

Proof of Concept

End-to-end execution was verified at commit 749af7b for representative cases per backend. Examples:

Numeric literal injection (all four backends). Feeding the following AST through compileToJavaScript emits a live console.log on call:

{ type: 'literal', value: "(console.log('PWN_INT_JS'),1)" }
// emits: (function(_) { return (console.log('PWN_INT_JS'),1); })(null);

The same AST against Ruby with "(system('touch /tmp/pwn_int_ruby'); 1)" creates the file; Python with "(print('PWN_INT_PY') or 1)" prints the sentinel; SQL emits the raw string into the numeric position.

SQL variable.name injection:

{ type: 'variable', name: "1; DROP TABLE users; --" }
// emits: 1; DROP TABLE users; --

Ruby datapath segment shell-exec:

{ type: 'datapath', segments: ["a, system(`touch /tmp/pwn_dp_ruby`)"] }
// emits: [:a, system(`touch /tmp/pwn_dp_ruby`)]

All ten end-to-end cases are automated in the ast-exec.js / ast-exec2.js scripts attached to the original report; each console.log and sentinel file was observed.

Independence from GHSA #2

This finding is independent of GHSA #2 (the .constructor.constructor sandbox escape). That report covers payloads delivered as plain Elo strings through the parser; the present advisory covers payloads delivered as AST objects through the programmatic API. The two threat models do not overlap: GHSA #2's identifier regex ^[a-zA-Z_][a-zA-Z0-9_]*$ accepts constructor (and is the right behaviour for a parser), while the present advisory adds runtime validation at every backend emit site so that AST-API consumers cannot inject arbitrary text into emitted source.