AI Diff Guard
Trained on the failure patterns of LLM-generated code

Did your AI assistant slip a bug into that PR?

AI Diff Guard flags AI-authored pull requests and runs differential analysis for the subtle bugs Claude, Copilot and Cursor commonly introduce — off-by-one, edge cases, swallowed errors — then posts a confidence-scored review. It’s not a generic linter.

Try it on a diff Add to a repo

Free for public repos · 20 PR reviews/mo · no card required · 5 diffs analysed

See it on your own diff

Paste a unified diff. Same engine the bot runs on every PR — no install.

Runs the same engine the GitHub bot uses. No sign-up, no install. Free analyses are rate-limited per month.

Your confidence-scored review appears here.

Try the sample — it has an off-by-one and an unguarded lookup.

How it works

Step 1

Install on a repo

Add the GitHub App. It listens for pull requests — no CI config, no setup.

Step 2

Flag AI authorship

Detects Claude / Copilot / Cursor / GPT commits from trailers, signatures and patterns.

Step 3

Differential analysis

Targets the specific bug classes LLMs introduce — not 500 generic lint warnings.

Step 4

Confidence-scored review

Posts a single review comment: risk score, findings, fixes. Signal, not noise.

The bugs we hunt

The specific defects models slip in even when the code compiles and looks right.

Off-by-one

Loop bounds, slice ranges, inclusive/exclusive confusion the model gets subtly wrong.

Edge cases

Empty input, single element, zero, negatives, unicode, timezones — the paths it never tested.

Swallowed errors

try/catch that hides the real failure, a missing await inside the try, no rollback.

Null / undefined

Optional chaining gaps and lookups assumed to hit — the classic ‘worked on the happy path’.

Silent fallbacks

Returns a default where it should raise — the dangerous ‘looks fine, hides corruption’ pattern.

Async races

Fire-and-forget promises, shared mutable state across concurrent calls, init races.

Not another linter

A linter flags style. AI Diff Guard reasons about behaviour — the way an LLM’s confident, plausible code hides a boundary error or a swallowed exception. Every finding comes with a confidence score so you can triage instantly, and a risk dashboard shows which AI-touched files break most.

Guard your repo

Confidence-scored

Triage by likelihood, not a wall of warnings.

AI-aware

Flags which PRs are AI-authored and weights the review.

Risk dashboard

See which AI-touched files keep breaking.

Be first when private-repo support ships

Drop your email — we’ll send the GitHub Marketplace link at launch.