Catalog
Each tool does one job. Here is what it does, how it works, and what it keeps. New tools are added here as they are ready.
AppGuard
AvailableSee what a stranger can see in your app.
AppGuard runs read-only tests on your live app, the same requests any visitor could make, and shows what is reachable. It is built for apps made with AI tools on Supabase.
How it works
- 1
Verify
Prove you own the site with one line of HTML or a small file.
- 2
Scan
We make a few hundred read-only requests at a gentle pace.
- 3
Prove
Each finding shows what was reachable, as counts and fingerprints, never your data.
- 4
Fix
Copy a ready-made prompt or SQL fix into your builder.
- 5
Test again
We run the same test and mark the finding fixed, or tell you what is left.
What you get
- Twelve checks, from open database tables to email records and certificates
- Every result labelled Proven, Likely or Info
- A fix you can paste, in plain language
- Daily re-checks once you turn monitoring on
What we keep
Counts, fingerprints and the first eight characters of a key. Never rows, files or full keys.
Good to know
It checks what is visible from outside. It does not read your code or log in as your users.
SwarmWatch
AvailableSpot groups of visitors acting together.
SwarmWatch looks at your web traffic for groups of visitors that behave almost identically, such as many addresses asking for the same unusual pages in the same half hour, and raises an alert you can investigate.
How it works
- 1
Connect
Link Vercel or Cloudflare, or send web server logs. Cloudflare Free plans use a small Worker.
- 2
Reduce
Each visit is reduced on arrival: no addresses, no browser text, no query strings.
- 3
Compare
Visits are held for up to an hour so groups can be seen across batches.
- 4
Alert
When a group looks coordinated, you get an alert with the pages and networks involved.
What you get
- Alerts that explain what was seen in plain words
- Mark each alert resolved or dismissed
- Works with the host you already use
What we keep
Alert summaries. The hour of reduced visits is then deleted.
Good to know
An alert is a lead, not proof of harm. Search engines and school networks can look like a group.
AgentGate
AvailableRules for what your AI agents may do.
Your AI agents ask AgentGate before they use a tool. Your rules decide: allow it, hold it for a person to approve, or block it. It also flags hidden instructions in outside content and agents that loop or run up a bill.
How it works
- 1
Write rules
List the tools an agent may use and what kind each is. Start from an example.
- 2
Try it
Pretend an agent is about to act and see what your rules would say. No code needed.
- 3
Connect
Add one small file to your agent (Python or JavaScript) and ask before each action.
- 4
Approve
Held actions wait for you. Approve or deny them in one click.
What you get
- Allow, hold or block, with the reason in plain words
- Limits on payments and on how often a tool is called
- A record of who asked for what, and what happened
What we keep
Which agent asked for which tool, the outcome and the names of the details sent. Never the values.
Good to know
Rules reduce risk but do not catch everything. Keep a person in the loop for actions that matter.
Fix drafts
Optional extraA drafted change for you to review.
For a proven finding, connect your GitHub repository and we draft a change on a separate branch and open a pull request. You review it. We never merge for you.
How it works
- 1
Connect
Install our GitHub App on one repository.
- 2
Draft
We read a few relevant files, with secrets removed, and propose a change.
- 3
Check
Our own rules reject anything risky before a pull request is opened.
- 4
Re-test
About ten minutes after you merge, we test again and mark the finding fixed if it is gone.
What you get
- A pull request ready to read
- Plain-language reasons for each change
- An automatic re-test
What we keep
The repository name and the outcome of each attempt. Never your code or the change.
Good to know
Your relevant files are sent to our AI model provider to draft the change. A person must review it.
War Room
ExperimentalAttack or defend an invented app in about seven minutes.
War Room is a short game for security researchers. You attack or defend an invented web app against the computer: twelve rounds, a few action points to plan each round, and no timer. It runs only on pretend apps and sends nothing to anything real. A short guided tutorial teaches each side. Matches against other players are being explored.
How it works
- 1
Pick a side
Attack to find what is exposed, or defend to fix it first.
- 2
Plan a round
Spend a few action points from a menu. Both sides plan in secret at the same moment.
- 3
See what happened
A defender sees alerts, and some are false alarms. An attacker sees only what it has found.
- 4
Replay
Every match can be watched back move by move.
What you get
- A daily game with the same app for everyone, and leaderboards by username for attackers and defenders
- A guided tutorial for each side that records nothing
- Opponents modelled on the visitors our research decoys actually see
What we keep
The plans you commit in every match, how long you took and the outcome, published only as totals. Groups under 5 players are hidden. The tutorial is not recorded.
Good to know
Experimental. Open to approved researchers for now, and it may change or be withdrawn.
Looking for our research?
We run decoy apps and an agent-versus-agent benchmark, and publish what we find. It is not a product. It is how we learn.
Read our research