Security for people who build with AI

Stop strangers from tricking your robots into stealing your keys.

When you build on GitHub, little helper robots run your project for you. TriDelPhi checks whether any of them could be fooled by a stranger into handing over your passwords and keys — and shows you the exact one to fix. Free, plain English, one line to turn on.

how-you-get-hacked.mov watch

GitHub runs little helper robots for your project — testing code, replying to comments, doing chores.

Animation paused — you have reduced motion on.

$ pipx install tridelphi $ tridelphi .

No account. No sign-up. Nothing leaves your computer. It just reads your files and tells you the truth.

Wait — what is this, really?

Your robots are helpful. That's exactly why they're worth tricking.

You don't need to be technical to get this. Here's the whole idea in three pictures.

You have helper robots

GitHub is where your project lives. It can run little robots that do jobs for you automatically — check your code, deploy your site, or answer comments with AI. Most people set these up without thinking twice.

Some robots hold your keys

To do their jobs, robots are handed your passwords, API keys, and permissions. That's normal and necessary — a robot that deploys your site needs the key to do it. The keys are the valuable part.

Anyone can talk to them

Here's the catch: a total stranger can leave a comment or suggest a change to your project. If a robot that holds your keys reads that stranger's words and follows them — your keys walk out the door. That's the whole danger, and it's more common than you'd think.

The one rule everything is built on

Two powers is fine. All three together is a disaster.

Every robot can have up to three powers. On their own, none is scary. The trap is a single robot holding all three at once — that's the recipe a stranger needs. Tap a power below to switch it off and watch the danger vanish.

Tap any circle to toggle that power on or off.

It reads words from strangers

Anyone on the internet can leave a comment or suggest a change — and that text reaches the robot.

It holds your keys

Your passwords, secret keys, or permission to change things. The prize a stranger is after.

It can reach the internet

So it can send things out — which is exactly how your keys would leave if it were tricked.

Any two of these is a normal, safe robot. Put all three in one robot and a stranger's comment becomes a way to walk off with your keys. TriDelPhi finds those robots.

One command · seven layers

A whole security team, bundled into one tool.

Keeping a project safe normally means juggling a pile of separate tools — a secret scanner here, a dependency checker there, a code auditor, a settings inspector — each with its own install, its own dashboard, its own jargon. TriDelPhi runs all of them for you, in order, and hands back one plain-English report. Start with a single check, or switch on all seven with one number.

separate tools, normallyEach one to find, install, learn, and read on its own.

command, hereOne line runs the whole ladder and merges every result.

and nothing leavesFree, open-source, no account — it reads your files and reports.

Find keys you left lying aroundScans your files for passwords and keys committed in the open.

replacesa secret scanner

Catch recalled building blocksFlags the outside code you use that's known to be vulnerable.

replacesa dependency scanner

Spot the robot with all three powersThe Rule-of-Two check no other tool does — TriDelPhi's own heart.

only herethe reason it exists

Grade your safety habitsScores the healthy settings a security review would check.

replacesa posture scanner

Proofread your code for dangerFinds the risky patterns that turn bugs into break-ins.

replacesa code auditor

Keep a signed receipt, and a bouncerProves what was checked, and can block a change that fails.

replacesattestation + a gate

Notice if a trusted tool gets swappedStops the build if an outside tool's owner or version changes on you.

replacessupply-chain monitoring

tridelphi . --level 1 → --level 7

Dial the coverage up or down with one number. Start with just secrets and the core check; climb to all seven when you're ready. Same tool, same one report — you decide how much.

The deep dive — every layer, in plain English

What each step does for you, step by step.

Here's every check up close — what it protects you from, told as a plain safety question, with the grown-up name and the open-source tool underneath in case you're curious.

The main event · Level 3

Is any one robot holding all three dangerous powers?

This is the check nothing else on the market does, and the reason TriDelPhi exists. It maps out every robot and asks the Rule-of-Two question: does one of them read stranger text and hold your keys and reach the internet — the exact recipe for theft? If so, it names the robot, shows the three powers, and tells you the single easiest thing to change.

In grown-up terms: a static capability-graph analysis of your GitHub Actions jobs, flagging any job where untrusted input, privilege, and egress intersect.

Level 1

Did you leave a key lying out in the open?

Like checking you didn't tape your house key to the front door. It scans your files for passwords and keys you accidentally left written in plain sight, where anyone who can see your code could grab them.

Grown-up name: secret scanning — we run gitleaks to find committed credentials.

Level 2

Are any of your ingredients on a recall list?

Your project is built partly from other people's code — like a recipe using store-bought ingredients. This checks whether any of those ingredients are known to be broken or dangerous, so you can swap them before they bite you.

Grown-up name: dependency scanning — osv-scanner matches your libraries against a database of known vulnerabilities.

Level 3 · also here

Are your robots set up safely to begin with?

Before the big Rule-of-Two check, this looks for the common sloppy setups — robots pulling in tools by a nickname that could be swapped out, or copy-pasting stranger text straight into a command. The everyday hygiene stuff.

Grown-up name: workflow linting — zizmor catches unpinned actions and template injection.

Level 4

Is your project following good safety habits overall?

Like a quick home-safety inspection — are the doors locked, is there a smoke detector? It scores your project on the healthy habits security folks look for, so you can see where you're strong and where you're exposed.

Grown-up name: posture checks — OSSF Scorecard grades branch protection, token permissions, and more.

Level 5

Does your actual app code have risky mistakes?

A careful proofreader for your code — not the robots, the app itself. It looks for the well-known dangerous patterns that let bugs turn into break-ins, and points at the exact lines.

Grown-up name: SAST (static application security testing) — powered by semgrep.

Level 6

Get a signed receipt, and a bouncer at the door.

Two things. First, it writes a tamper-proof receipt of exactly what was checked and signs it — proof you can show anyone. Second, it acts as a bouncer: if something failed the check, it can stop the line and refuse to let the change through.

Grown-up name: attest & gate — a signed in-toto evidence statement, plus policy enforcement as its own step.

Level 7

Would you notice if a stranger replaced your babysitter?

You trust certain outside tools. This memorizes exactly who they are today. If one quietly gets handed to a different owner — or swapped for a different version behind your back — it stops everything and tells you. This catches attacks that even careful people miss.

Grown-up name: a trust-lock — detects a signer or commit change in a dependency you consume, the takeover class that pinning alone can't see.

Scary words, decoded

Every bit of jargon on this page, in one place.

Security people love their vocabulary. You don't need it — but here's the translation so nothing on this page (or in a scary tweet) can confuse you again.

CI / "Actions" continuous integration

The helper robots GitHub runs for you automatically whenever you change your project — testing, building, deploying, answering comments.

RCE remote code execution

A stranger getting to run their own commands on your stuff. The nightmare outcome — it's what "getting hacked" usually means underneath.

Secrets

Your saved passwords and keys — the ones robots use to deploy your site, post to services, or spend money on your behalf.

Pull request "PR"

A suggested change to your code. Helpful when it's from a teammate — but anyone on the internet can send one, which is where the danger sneaks in.

Prompt injection

Hiding secret orders inside text your AI reads. The AI can't always tell "content" from "commands," so buried instructions can hijack it.

SARIF

A standard report format GitHub understands. It just means the findings show up neatly in the place you already look, instead of a wall of text.

Turn it on

One line, and it watches every change for you.

You don't need to learn any of the checks or flags. Run one command and TriDelPhi sets up a robot of its own — a good one — that reads every suggested change and leaves a plain-English note if it spots trouble.

your terminal

# 1. install it (once)
$ pipx install tridelphi

# 2. set up the guard robot
$ tridelphi init
✓ wrote .github/workflows/tridelphi.yml

# 3. commit & push. that's it —
#    it now checks every change and
#    leaves a plain-English note.

Three steps, then never think about it again.

The guard robot runs on GitHub, for free. On every suggested change it posts a friendly comment: all clear, or here's the one thing to fix and how. No dashboards, no subscription, no data leaving your project.

Install itOne line. Works with pip or uvx too — Python is all it needs.

Let it set up the guardtridelphi init writes the robot for you. Nothing to configure.

Commit and forgetIt comments on every change from then on. Green means you're good.

free & open source no account or token nothing leaves your machine already using a workflow? one uses: line

Built on the security world's best open tools — and it says so.

TriDelPhi doesn't reinvent the wheel. It runs the tools professionals already trust and stitches their results into one plain-English report. Every one is credited, by name, with its license.

gitleaks osv-scanner zizmor OSSF Scorecard semgrep