10 Break

We break agent companies on purpose.

An agent company is more than a model: agents, the tools they hold, the screens they drive, the gates that wait for a person, and the money they can spend. We attack our own on a schedule to find where it fails. With your written go-ahead, we do the same to yours, then show you how to fix what we find.

01 · What we attack

Everything your agents can read, touch or spend.

We test the agents you have deployed, as they run, with the access they really hold.

Support and sales agents

Agents that talk to your customers in chat, mail or social replies, and what a stranger can make them say or do.

Internal agents

Agents that read your mail, tickets, documents and code, and act on them: what reaches them from outside, and where it goes next.

Tool servers

The tools and servers your agents call: what each one lets an agent do, and whether a tool can be talked into more.

Browser and phone agents

Agents that drive a screen. A page, a pop-up or a message can steer them wrong, or let them press what they were never meant to.

How they fail

  • Instructions hidden in mail, web pages, documents and social posts that an agent reads as orders.
  • Agents getting past an approval gate: sending, merging, paying or deleting without the person who should decide.
  • Tools used for what they were not given: a read tool that writes, a search that reaches private data.
  • Keys and private data leaking into replies, logs, drafts or another tool’s input.
  • Runaway spend: loops, retries and fan-out that burn tokens, machines or money before anyone looks.

02 · How an engagement runs

Scope, permission, attack, report, retest.

Nothing is tested before both sides have signed what is in scope. Every finding comes back with the steps to reproduce it and a fix you can ship.

  1. 01

    Scope

    Which agents, tools, accounts and environments are in, which are out, and what would count as harm.

  2. 02

    Permission in writing

    Signed by someone who owns the systems. It names the targets, the dates, the methods and who to call to stop.

  3. 03

    Attack

    We test only what the paper names, from the accounts and networks it names. A live risk to your data or customers stops the test, and you hear the same day.

  4. 04

    Report

    Every failure with its reproduction, what it could cost you, and a fix: a gate, a tool change, a prompt change or a check.

  5. 05

    Retest

    After you ship the fixes we run the same attacks again and say what still breaks.

Ongoing, if you want it

The attacks that worked keep running against your agents on a schedule, and each new failure goes straight to the people who own the defense. Break feeds Guard.

03 · On the record

Our own failures, and what we changed.

We run a company on agents, so we break first. These are real failures in our own company, written without secrets, addresses or names. Each tag reads from the job that broke to the job that holds the fix.

  1. One unclear post stopped every account for thirty hours.

    What broke
    A worker phone posted a story, then checked it by opening the newest frame. The app showed an older one, so the post was marked unclear, and the safety breaker shut posting for every account until a person looked.
    How we found it
    Accounts went quiet. Reading the worker’s own screenshots showed the post had gone out.
    What we changed
    A post now counts when the account’s frame count grows by exactly the number it sent. A watchdog closes the breaker on that evidence, resets it after fifteen minutes with an alert, and stays shut if three posts are unclear within six hours.
  2. A mail sync fanned out and choked a shared machine.

    What broke
    The first sync of two mailboxes started one process per message, six at a time per mailbox, on a four-core machine that also serves our code. Load reached nineteen and remote logins began to drop.
    How we found it
    The machine stopped answering logins. We traced it to the sync and stopped it within twelve minutes.
    What we changed
    Calls are batched under a hard process budget, the first sync reads a window at a time and can pause, and per-item work never fans out against a shared machine.

    Lumbridge Mail

  3. Two hands on one phone, and a scheduled post failed.

    What broke
    An agent used a worker phone for a lookup while a scheduled post was due on the same phone. The post failed halfway through, and other worker phones stalled mid-task on screens they did not expect.
    How we found it
    The scheduled job failed at a tap, and its log showed someone else had moved the screen.
    What we changed
    One primary worker phone takes the queue. Anyone using it by hand checks the queue first and stays clear of scheduled slots. Every phone job ends as done, stuck or failed, with loop detection, so a stuck job says so instead of tapping forever.
  4. One wrong request benched the only reader account.

    What broke
    A new lookup tool asked for search results the way the service refuses. The error put our one shared reader session in a cooldown, and the news feed that depends on it stalled.
    How we found it
    Testing the lookups, we saw the news feed stop on the same minute as each error.
    What we changed
    Lookups now ask the way the service accepts, in their own process. They hand the session back without a cooldown, never retry a refusal, and share a cached hourly budget with the feed.
  5. A stranger’s words can’t give our agents orders.

    What broke
    Our assistant reads mail, web pages, screens, captions and podcast transcripts all day. Any of them can contain a line written to look like an instruction from us.
    How we found it
    Every place a stranger can write is treated as data, never instructions.
    What we changed
    Everything a stranger wrote reaches an agent labeled as data. Mail that asks to forward, reply with a code or ignore the rules is flagged to a person and nothing else happens. The assistant has no send tool, and a line in a transcript can’t approve a post.

Engagements for other companies stay confidential. What we publish here is our own.

04 · Open evaluations

Every failure we can scrub becomes a test anyone can run.

With your consent, an attack that worked is stripped of your names, data and systems and turned into an open evaluation: a small agent company with the same weakness, and a check that says whether an agent falls for it.

Scrubbed first

Nothing identifying leaves the engagement. You see the scrubbed case before it is published, and you can say no.

Run in the open

The evaluations run on compute that people and labs contribute to open work, and the results are published.

Better defaults

What the evaluations teach goes back into the engine’s own gates, so every Lumbridge install gets the fix.

Contribute compute

05 · The rules

Authorized work only.

  • We test only systems we were asked to test, by someone with the right to ask, in writing.
  • We never attack a system, account or person outside the agreed scope, including the services your agents call.
  • We work on your agents and your setup. We don’t sell ways around other companies’ models.
  • Your data stays in the engagement: used to find and fix failures, deleted at the end unless you ask us to keep it.
  • Either side can stop the work at any time, and a live risk stops it at once.

06 · Straight answers

What people ask us.

Do you need our agents to run on Lumbridge?

No. We test the agents you have, wherever they run and whatever they are built on.

Do you test the model itself?

We test what your agents do with it: what they read, the tools they hold, the gates around them. The model is one part of that.

Will testing touch our customers?

Only if the scope says so. Most engagements run against a copy, a test account or a quiet hour, and the paper says which.

Who sees the report?

You do. It is written for the engineers who will fix it and for the person who signed the scope.

How much does it cost?

It depends on how many agents, tools and environments are in scope. Write to us and we will answer with a scope and a price.

07 · Write to us

Tell us what your agents do.

This page sends nothing. The button opens an email in your own mail app with a few questions to answer; a person reads every one.

To: hello@lumbridgecorp.com
Subject: Break: testing our agents

Hello Lumbridge,

Company:
The agents we want tested (support, internal, tools, browser or phone):
What they can read, change or spend:
Where they run:
Who can sign the scope:
When we would like to start:

Written on lumbridgecorp.com/break

Write the emailOr write to hello@lumbridgecorp.com

Want the defense built too?

We set up Lumbridge on your machines with the gates already in place, and stay accountable for it.