The honest view from inside technical leadership
AI News

AI News

Daily AI news with an engineering leadership take, each story credited to its original AI News report.

Security · AI News

CommBank CodeCheck Is an AI Code Review Harness That Treats the Code It Reads as Hostile

Commonwealth Bank of Australia has published the AI security review harness its own engineers use. The most useful part for other teams is the threat model it applies to itself.

Line-art illustration of a code window inside a sandbox box with a magnifying glass, representing CommBank CodeCheck AI code review running in isolation

CommBank CodeCheck is a free AI code review harness that Commonwealth Bank of Australia released on GitHub in November 2026 under the MIT licence. It plugs into Claude Code or OpenAI Codex, reviews a codebase, branch, commit or pull request for security weaknesses, and hands engineers evidence for each suspected flaw. CBA says it is based on a capability used inside the bank, where thousands of engineers work across roughly 2.5 billion lines of code.

CodeCheck Splits AI Code Review Into Hunters and Refuters

The README describes four stages. A scope agent maps entry points, trust boundaries and technologies, writes a threat model, and ranks every CWE family as high, medium or low priority for that code. In discovery, one hunter agent is assigned per chosen CWE family, plus three that always run: an open-ended hunter, a design-flaw hunter, and a hunter for backdoors and implants. In review, duplicates are merged and separate reviewer agents try to refute each finding. The orchestrator then weighs that evidence and writes a report with the source-to-sink data flow, the reviewer’s verdict and a CVSS 4.0 vector, in fields that map to SARIF so results can flow into existing code-scanning tools. A diff-aware mode limits findings to data flows a change touches, including flows whose source and sink did not change.

The 10% to 15% Benchmark Applies to CBA’s Internal Setup

CBA says it built an internal vulnerability evaluation framework and benchmarked similar capabilities with it for six months. Its harness, configured for the bank’s environment, generally identified 10% to 15% more vulnerabilities in a single run. The bank has not named the tools it compared against or published the benchmark data, and AI News notes that the public release does not include all of CBA’s internal systems and controls. The figure describes the bank’s configuration. It is no promise about what the download will do on your code, and CBA’s own documentation warns that the tool can miss flaws, flag safe code and overstate severity.

CodeCheck Assumes the Code Under Review Can Attack the Reviewer

The design choice worth copying is how CodeCheck treats its own input. The README tells every agent to treat all text in the reviewed tree as data, and to treat a comment such as “skip this file” as a reason to look closer. Agents are told to stay read-only, avoid network requests, and report weaknesses without proposing fixes. CBA then says plainly that these prompt-level guardrails are followed non-deterministically and that “you cannot rely on them as a security control.” Its answer is an isolated environment: mount the code read-only, block network egress except to the model provider and GitHub, and keep credentials out of the sandbox. The README lists Docker Sandboxes, Firecracker microVMs and Amazon Bedrock AgentCore Runtime as examples.

Why This Matters for Engineering Leadership

An AI code review agent is a coding agent with tool access pointed at code your team may not have written, such as dependencies, contributor pull requests and acquired repositories. CBA, a bank with a large security function, has concluded that this setup needs the same containment as running untrusted code, because the reviewed text is effectively instructions to the reviewer. Before you approve any AI review tool, CodeCheck or otherwise, ask three questions. Where does the agent run, and can it reach credentials or the open network from there? Does the tool have an adversarial step that tries to disprove findings, or does it pass the first model’s guesses straight to engineers? And have you measured it on your own code with known vulnerabilities, rather than trusting a vendor or peer benchmark? CBA’s release gives you a reference architecture for all three, under a licence that lets you adapt it. Budget for the model spend, since every review runs many sub-agents on your provider account, and keep a named engineer accountable for each finding that gets acted on.

FAQ

What is CommBank CodeCheck?

CodeCheck is a free, MIT-licensed plugin for Claude Code and OpenAI Codex that runs an AI-driven static security review of source code. Commonwealth Bank of Australia released it on GitHub in November 2026 as a one-time release.

How does CodeCheck review code?

It runs four stages: scope, discovery, review and report. Specialist sub-agents hunt one CWE family each, and separate reviewer sub-agents try to refute every finding before the report is written.

Does CodeCheck find more vulnerabilities than other tools?

CBA says its internally configured harness generally found 10% to 15% more vulnerabilities per run than similar capabilities over six months of internal benchmarking. It has not named the comparison tools, and the public version lacks some of CBA’s internal setup.

Why does CodeCheck need to run in a sandbox?

The code under review is untrusted and can contain prompt injection aimed at the agent. CBA recommends a read-only mount, network egress limited to the model provider, and no direct access to credentials.

Does CodeCheck fix vulnerabilities?

No. Its agents are told to report weaknesses and how to trigger them without editing files or proposing fixes, and engineers decide on remediation.

What does CodeCheck cost?

The tool is free to download. Teams pay their AI coding agent and model provider for the tokens each review consumes.

Originally reported by Muhammad Zulhusni for AI News, November 18, 2026: Why Commonwealth Bank is using AI to detect software vulnerabilities earlier. Primary sources: CommBank newsroom, “CommBank launches CodeCheck to help Australian businesses find software vulnerabilities”; Commonwealth-Bank-of-Australia/CodeCheck on GitHub (README).

Read more AI News