tethercode.
Get early access

tethercode / access and confidentiality

Access control for AI agents: who may see what, what may leave, what may be done

Access is given by data source, not by rank. Every model call is read before it is sent, by a small model trained for that one job. A risky action waits for the person whose agent it is.

Nothing confidential leaves by accident

Agents read things: a tool result, a customer record, a settings file, somebody's chat export. Every read changes what the agent is holding, and so what it should be allowed to do next. An agent that has just opened a customer's account should not be sending a summary to a public model thirty seconds later. Nobody decides to leak. It happens one helpful step at a time, and the usual safeguards do not hold.

Access handed out by rank

"Senior staff may see confidential material" sounds careful. In practice one permission opens every confidential thing in the company, and every agent that person starts carries all of it.

Here: access is given one data source at a time. Access to payroll opens payroll and nothing else. An agent acts as the person who started it, with that person's access and no more, and no administrator can read everything.

The sender says how sensitive it is

Many systems let the application, or the agent itself, state what it is sending. That is asking the fox whether it may enter the hen house. One fixed setting per project is no better: a project holds public notes and customer records side by side.

Here: nobody states anything. Every model call is read at the moment it is made, by a small model trained for that one job, and it travels only as far as your own rules allow for what is actually in it.

A list of dangerous commands

A list catches the delete it was written for. It cannot see an insulting email to the whole company, a post that gives away next year's prices, or a delete spelled in a way nobody thought of.

Here: the list stays, and a second trained check judges what the action would do: lose data, take a service down, hurt someone, embarrass the company. If it could, the action waits for a person.

A general model as the judge

A general-purpose model has never seen where your company draws the line between internal and public. Asked to judge, it answers with confidence and is often wrong in the direction that hurts.

Here: the judge is called Kev: a small model with a part trained separately for each of these two decisions, and built to err towards caution. The numbers are below.

Why we trained our own judge

Each dot is one of 1,000 test samples of company-style text, 250 at each of four sensitivity levels. Two judges saw the same samples: Kev, and Jev, a general-purpose hosted decision model. The colour of a dot says what the judge did with that sample.

  • marked public when it was not
  • rated too low in another way
  • rated too carefully
  • right

Read it by colour. Red is content that could go anywhere when it should not. Yellow is content rated a level or two too low. Blue is the harmless mistake, too careful. Kev's field has no red and four yellow dots. Nearly half of everything the general model called public was not: 159 of 347.

The same for actions

Kev's second job is judging what an agent is about to do. Each dot is one risky action. Green was stopped before it ran.

Before training, the same model could only catch as many by stopping 96 to 100 of every 100 actions, safe or not. The price of the trained one: about one safe action in six also stops to ask, closer to one in four on raw shell commands. Each is one tap on your phone.

Fast enough to sit in front of every call

Kev
39 ms
Jev
418 ms

Slowest of 1,000: 178 ms against 4,014 ms. Most of the gap is the network: Kev ran on a graphics card in the room, Jev over the internet.

Read these the right way

The samples are written to resemble company text from twelve industries, with deliberately hard cases. We do not train a secrets detector on real secrets. Real documents are messier, so this is the best case, not a promise.

Jev was judged on our definitions, which it had never seen. This shows what training for the job buys, not which model is better.

Zero is what we measured, not a guarantee: 0 wrong in 215 still leaves room for a true error rate of up to 2 in 100. The classifier in this comparison is our first generation; the one that ships today caught 99.86% of the most sensitive samples on the full test of 8,148, against 99.73%. Action labels were set by a larger model and have not yet been checked by people. The method and the full tables.

Too careful costs a tap. Too trusting is the incident.

Who may see what

Access is given to a person, one data source at a time. Their agents get exactly that, and nobody gets everything.

Asked: why can't my agent open the payroll export?
Who
Dana, through her agent Codex
Wants
payroll exports, a confidential source
Holds here
nothing: no role or grant of hers names this source
Holds elsewhere
finance reports and the billing code through the Finance team
So
refused, and told why

The answer names the role or grant behind every access a person has.

For you

One source at a time

A grant names one data source: a repository, a set of documents, a company database. It can come from a role, a team, a direct grant, an approved request or ownership. Holding it is the permission. Nothing else is.

Four classes, one of them open

You class each source as public, internal, confidential or personal data. Public is readable by everyone in the company. The other three are readable only by people who hold a grant on that source.

Nobody sees everything

There are no clearance levels and no administrator override. Managing access is not the same as having it.

See why, and see what a change would do

For any person, tethercode shows which role or grant gives them an access, or that none does. Before you change a role, it shows who the change would affect.

Your sign-in, your groups Enterprise

People sign in through the company sign-in, and roles follow your directory's groups. When a person leaves, their access ends, and their agents and open tasks are frozen and kept, not lost.

For your agents

An agent is its person

An agent acts for the person who started it, with that person's access to each source and nothing more. Every action carries that person's name.

Told what was withheld

When a search touches something its person may not read, the agent is told that something was withheld and how much. It is not shown it.

Private memory stays private

A person's own notes and their agents' memory are a source like any other, owned by that person. A colleague reads them only if the owner shares them.

Agents never hold your secrets

Passwords and keys live in a vault. The service that needs one borrows it for a moment at the point of use, every read is checked and recorded, and changing a secret keeps the old version until the new one is seen to work.

A record nobody can rewrite

Access decisions are recorded. Records can only be added, even by the owner of the database, and an action whose record cannot be written does not happen.

What may leave

Every model call tethercode makes is read at the moment it is made, and goes only as far as your rules allow for what is actually in it.

One model call, read before it goes
Asked
summarise this week's open tickets for the Acme account
Carries
four support tickets and two rows from the customers table
Kev reads it as
confidential
Your own label
the customers table is marked most sensitive
Treated as
most sensitive: the strictest reading wins
May reach
models on your own network only
Sent to
your local model

The record keeps both readings and which one decided.

For you

Four levels, four places, your map

Content is one of four levels, from public to most sensitive. A model sits in one of four places: your own network, your private cloud, a provider you have a contract with, or the public cloud. You set how far each level may travel. A bank keeps everything at home; a small team may allow more.

Your own labels count

Mark a table, a column or a document as more sensitive than the place it sits in, and that label travels with every row into every call. A label can only raise the level, never lower it.

The owner of the data has the last word

A data source can carry its own rule: never to a model outside this machine, or never outside the company. The rule goes wherever the data goes.

It fails towards private

If Kev cannot be reached, the call is treated as the most sensitive kind and stays on your most private machines. It is not waved through and it is not refused, and being busy never widens where a call may go.

Where Kev itself runs

On Enterprise, on your own GPU: nothing is sent anywhere to be judged. On Free and Premium it is one of the hosted helper models, so the text of a call goes to the provider you chose for it, to be read.

Where it is too careful today

Kev rates very short, chatty messages as most sensitive. That costs you model choice on small talk, not privacy. The fix is more training on short text, and it is on our list.

For your agents

Nobody declares anything

Neither the agent nor the application says how sensitive a call is. A caller can make a call stricter and can never make it looser.

All of it is read

The whole call is read: the messages, the tool results, the instructions. Text too long for one reading is split, every part is read, and the strictest part decides. Nothing is cut short to save time.

A picture is not judged by its caption

An image that nobody has classified makes the call most sensitive, whatever the text beside it says.

What is made from it stays marked

Stored content carries a sensitivity mark, and a summary, a report or a memory made from it inherits the strictest mark among its sources.

Every reading on the record

Each call's record keeps Kev's full answer, your labels and the stored marks, and says which of them decided. When someone asks why a call stayed local, there is an answer.

What may be done

Before an agent acts, the action is judged by what it would do. If it could cause harm, it waits for a person.

An action a list would not have caught
Agent
Codex, working for Dana
Wants to
post to the company account: "our Enterprise price goes up 20% in March"
The list
nothing: it is not a delete, a stop or a restart
Kev
could harm the company's reputation
needs Dana Codex wants to publish a post. Kev flagged it: reputational harm. ApproveDenyasked on Telegram

For you

Approve from your phone

A risky action stops and asks on Telegram or Slack, with buttons. The request says what the agent wants to do and why it was stopped.

Five questions about every action

Could it lose data? Take a service down? Hurt someone? Embarrass the company? And how serious is it overall? One yes that is strong enough, and the action waits.

The person whose agent it is decides

Each person approves their own agent's requests. There is no queue in front of an administrator. Only changes to secrets need a second person, on Enterprise.

Every decision recorded

What was asked, why, who answered and when. An approval covers the exact action it was given for, not the next one that looks like it.

For your agents

Two checks, either one is enough

A list of destructive commands, and Kev's judgement of the effect. An agent that tries "start" because "restart" was on the list still has to ask.

It sees what a list cannot

Sending, posting and publishing tools have no dangerous words in them. They are judged like everything else.

A long command is read whole

It is split into parts that join back to the exact original, each part is judged, and the riskiest part decides.

When Kev cannot answer, the list still stands

A failure is never taken as "allow" and never as "deny". The list decides alone, exactly as it would have without Kev.

More of tethercode

Try it on your own repository.

Free for 3 repositories, with no time limit and no card. One email when it is ready, and nothing else.

Deploying for a company? Talk to us. Prices are on the editions table.