Eval показывает 96% качества, тесты проходят, а после релиза LLM‑ассистент допускает критическую ошибку. Проблема может скрываться в самой системе оценки, которая пропускает опасные ответы и создаёт ложное ощущение надёжности. Разбираем семь ошибок тестирования LLM‑функций и способы построить eval, результатам которого действительно можно доверять. Проверить свой eval
All versions of package static-eval are vulnerable to Arbitrary Code Execution using FunctionExpressions and TemplateLiterals. PoC: var evaluate = require('static-eval'); var parse = require('esprima').parse; var src="(function (x) { return ${eval("console.log(global.process.mainModule.constructor._load('child_process').execSync('ls').toString())")} })()" var ast = parse(src).body[0].expression; evaluate(ast)
23andMe Yamale before 3.0.8 allows remote attackers to execute arbitrary code via a crafted schema file. The schema parser uses eval as part of its processing, and tries to protect from malicious expressions by limiting the builtins that are passed to the eval. When processing the schema, each line is run through Python's eval function to make the validator available. A well-constructed string within the schema rules can execute system commands; thus, by exploiting the vulnerability, an attacker can run…
Electron is a framework which lets you write cross-platform desktop applications using JavaScript, HTML and CSS. A Content-Security-Policy that disables eval, specifically setting a `script-src` directive and _not_ providing `unsafe-eval` in that directive, is not respected in renderers that have sandbox disabled. i.e. `sandbox: false` in the `webPreferences` object. This allows usage of methods like `eval()` and `new Function` unexpectedly which can result in an expanded attack surface. This issue only…