Evaluates a Python expression with ambient globals and locals, making input-to-code transitions difficult to constrain.
The pack does not provide a callable signature for this record; the watched access path is shown directly.
| Arg | Access path | Purpose | Watched |
|---|---|---|---|
| Argument[0] | Argument[0] | The access path Atropos marks for this model. | ▲ sink |
Attacker's-eye view · code-injection
An attacker supplies syntax that is evaluated with the process's privileges.
// modeled boundary: Argument[0]
eval(attacker_influenced_value);
1Keep code and data separate.
2Replace evaluation with a parser or allow-list.
3Constrain the execution context if evaluation remains.
These checks require the surrounding codebase. The model names the boundary; it does not decide reachability or prove that a guard dominates every path.
Is this a bug in your code?Atropos stops here — by design
Atropos identifies Argument[0] as a code-injection sink. It cannot see whether untrusted data reaches this call in your repository.
Lachesis is the codebase-level step: it traces reachability and guards for this symbol.
Check this symbol in Lachesis →Reviewed guidance
Safer direction: Parse structured data or expose a small allow-listed operation set instead of evaluating source text.
Unsafe shape
eval(user_expression);
Safer shape
operator.add(left, right);
Model facts · verbatim from the pack
| Role | Kind | Access path | Model ID | Confidence |
|---|---|---|---|---|
| sink | code-injection | Argument[0] | python.builtins.eval.a0 | high · corrob. 3 |