AI security review

Verifies

Requires an AI engine: Three-phase model review (repository context, comparison against the established security patterns, exploitability assessment) of authentication, authorization, input handling, crypto and code-execution paths, against a fixed taxonomy of ten categories each mapped to a CWE.

Inputs

Up to 40 security-relevant code files selected by path and sorted, each numbered by line and split into chunks of at most 16000 characters (first 160000 characters of a file), sent to the chosen engine through the platform ai service behind a stable repository-context prefix (languages, manifests, security-relevant dependencies, files under review). One execution spends at most 220000 estimated tokens; files that do not fit are named in the trace and not reviewed. Paths listed in excludes (test, spec and tmp files, __tests__, __mocks__, fixtures, test, tests and audit-scripts folders) are never analysed. Every model call (rubric, repository context and chunk) stays under 120000 characters. Speed pass (1.1.4): the chunks are sent in parallel up to the engine's concurrent-call limit and processed in plan order.

Decision rule

A finding counts only when it cites a line inside the reviewed chunk, carries a severity of HIGH, MEDIUM or LOW, a category of the taxonomy and a confidence of at least 0.8; findings matching the hard exclusions (denial of service, rate limiting, memory and resource leaks outside C or C++, regular-expression denial of service, SSRF inside HTML, documentation files, speculative wording) are discarded after the engine answers. FAIL on at least one HIGH finding; MEDIUM findings are blocking only when the script runs inside an Extended suite (suite id ending in -extended); LOW findings are informative. FAIL also when the engine answer for any reviewed chunk is empty or not parseable JSON: that file counts as not evaluated and a run with unevaluated files never passes. Without an AI engine the check is not executed (requires an AI engine).

Type

ai-assisted · engine chosen by the organisation · medium

1.1.4: SPEED PASS: the chunks of the plan go to the model IN PARALLEL up to what the engine admits (ctx.ai.maxConcurrentCalls, 1 when the sandbox does not publish it) instead of one after another, and the answers are processed in plan order, so the verdict and the trace never depend on which answer arrives first. Measured on the attestation of social: 40 sequential calls of ~4 000 tokens with a four-token answer, the model busy well under a second per call and the rest orchestration. Selection, rubric, taxonomy, exclusions and decision rule unchanged. (1.1.3: requires an AI engine (spec.requiresEngine: true): the check needs judgement and is executed only when the organisation provides an engine; the core catalogue runs without one. (1.1.2: never analyses temporary files (*.tmp.*; spec.excludes follows AUDITOR_ANALYSIS_EXCLUDES). (1.1.1: never analyses test, spec, fixture and mock paths nor the auditor's own scripts (spec.excludes = AUDITOR_ANALYSIS_EXCLUDES); a fixture-looking secret (sk_test_, example, dummy) in real code is LOW, informative; every model call stays under the engine prompt cap (chunks shrink to what the rubric and the repository context leave). (1.1.0 adopted the three-phase review, exclusions and confidence threshold, method from anthropics/claude-code-security-review, MIT.))))

The full script is disclosed on request in a read-only viewer (never published on GitHub); the attestation binds to this exact hash.