Mistakes in the dangerous direction
Rated less sensitive than it is.
- 4
- Kev
- 193
- Jev
Kev's four each slipped one level. Its other 50 mistakes went the harmless way: too careful.
tethercode / benchmark
An AI agent reads customer records, contracts and secrets as part of ordinary work. Something has to decide, on every call, whether what it is holding may leave the building. We gave that decision to two judges on the same 1,000 samples. One marked 159 non-public samples as public. The other marked none.
The same 1,000 samples of company-style text, 250 at each of the four sensitivity levels, given to two judges: Kev, trained for this job, and Jev, a general-purpose hosted decision model. The colour of a dot says what the judge did with that sample.
Kev ours, trained for this job
Most sensitive
Confidential
Internal
Public
Jev a general-purpose hosted model
Most sensitive
Confidential
Internal
Public
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 called 215 samples public, and all 215 were. Nearly half of everything the general model called public was not: 159 of 347, of which 18 were the most sensitive kind, 49 confidential and 92 internal.
Rated less sensitive than it is.
Kev's four each slipped one level. Its other 50 mistakes went the harmless way: too careful.
Green is Kev, grey is Jev. Kev's weakest row is public: it treated 35 public samples as more sensitive. That is the safe side to be wrong on.
Most of that gap is the network: Kev ran on a graphics card in the room, Jev over the internet. What it shows is that the check can sit in front of every call unnoticed.
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.
Public test set 515 of 527
Real commands it had never seen 58 of 59
The price: 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.
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 samples are synthetic and Jev was judged on definitions it had never seen; the method, the full tables and every limit are below.
Kev is the small model inside tethercode that reads every model call before it is sent and every action an agent is about to take. Below the charts are the method and the full tables behind them and behind the figures on the home page, including what they do not show.
The two ways of being wrong are not the same mistake. Calling a marketing page confidential costs somebody twenty seconds. Calling a customer record public puts it in a log, a search index or a summary sent to an outside model. One is friction. The other is the incident.
A plain accuracy score treats the two as equal. So the tables below separate them: how often each judge rated something as less sensitive than it is, and how often as more.
1,000 samples, 250 at each of four levels. Each sample is one piece of text an agent might read: a document, a chat thread, a tool's raw answer, a code file, a settings file, a table, a support ticket, an email thread or the record of an agent's own steps.
| Level | What is in it |
|---|---|
| Most sensitive | customer account and call records, personal data, health data, card data, core trade secrets, passwords and keys |
| Confidential | internal strategy, legal and personnel matters, contracts with vendors and partners |
| Internal | procedures, internal messages, who reports to whom |
| Public | marketing, company information, public documentation |
The samples are written by a generator, not taken from real companies: twelve industries, fifteen kinds of content and nine formats. We do not train or test a secrets detector on real secrets, and anything that matched a live key format or a valid card number was thrown out.
Easy examples make a classifier that fails on real work, so five kinds of hard case are mixed in:
Kev, in this comparison, is our first generation: an open model of 1.5 billion parameters with a small trained add-on, about 17 megabytes. It was trained on one rented graphics card in sixteen hours, for under twenty dollars. Under-rating was punished up to eight times harder than over-rating during training, most heavily for the most sensitive level.
Jev is a general-purpose hosted decision model, called through its public interface and given a written definition of each of the four levels.
Both saw the same 1,000 samples, once.
| Of 250 samples at each level | Kev found | Jev found |
|---|---|---|
| Most sensitive | 248 | 198 |
| Confidential | 235 | 173 |
| Internal | 248 | 122 |
| Public | 215 | 188 |
The other side of the same question: when a judge names a level, how often is it right?
| When it said | Kev was right | Jev was right |
|---|---|---|
| Most sensitive | 248 of 263 | 198 of 230 |
| Confidential | 235 of 264 | 173 of 292 |
| Internal | 248 of 258 | 122 of 131 |
| Public | 215 of 215 | 188 of 347 |
The last row is the one that matters most. "Public" is the label that lets content go anywhere. Kev used it 215 times and was right every time. The general model used it 347 times, and 159 of those were not public: 18 were the most sensitive kind, 49 confidential and 92 internal.
Kev's weak side shows in the same tables: it found only 215 of the 250 public samples, because it treated 35 of them as more sensitive than they are. That is the direction we chose for it to be wrong in.
| Mistakes | Kev | Jev |
|---|---|---|
| Rated less sensitive than it is | 4 | 193 |
| of which one level too low | 4 | 118 |
| of which two levels too low | 0 | 57 |
| of which three levels too low, most sensitive to public | 0 | 18 |
| Rated more sensitive than it is | 50 | 126 |
| All mistakes | 54 | 319 |
Where the two judges disagreed and one of them was right, Kev was right 286 times and the general model 21 times.
Kev also knows when it is sure. It gave 698 of the 1,000 samples a confidence of 99% or more, and all 698 were right. Below that, its answer is better treated as a request for a second look than as a verdict. The general model's stated confidence did not track how often it was right.
| Time to judge one sample | Kev | Jev |
|---|---|---|
| Average | 38.7 ms | 418.0 ms |
| Middle sample | 38.0 ms | 403.2 ms |
| Slower than 95 in 100 | 42.0 ms | 495.3 ms |
| Slower than 99 in 100 | 45.0 ms | 593.7 ms |
| Slowest of 1,000 | 178 ms | 4,014 ms |
Most of this gap is the network, not the model. Kev ran on a graphics card in the same room; the general model was reached over the internet. A general model hosted locally would close much of the average. What the table does show is that a check like this can sit in front of every single model call without anyone noticing, and that a four-second worst case is what an agent's loop would feel.
A list of dangerous commands catches the delete it was written for. It cannot see an insulting email to the whole company or a post that gives away next year's prices. So Kev has a second trained part that reads one action an agent is about to take and answers five questions: could it lose data, take a service down, hurt someone, embarrass the company, and how serious is it overall. If any answer is strong enough, the action waits for a person.
It was trained on about 21,000 labelled actions drawn from public agent-safety test sets, a large collection of ordinary shell commands, and real commands from our own machines. No command appears in both the training set and a test set.
| At the setting we ship | Risky actions caught | Safe actions that also asked |
|---|---|---|
| Public test set, 2,417 actions | 515 of 527 | 331 of 1,890 (17.5%) |
| Real commands it had never seen, 938 actions | 58 of 59 | 206 of 879 (23.4%) |
The same model before this training could not be used as a gate at all. To catch 95 in 100 risky actions it had to stop between 96 and 100 of every 100 actions, safe or not.
In a live check of hand-picked actions, it stopped all seven risky ones, each for the right reason: deleting a database's data folder, shutting down services together with their stored data, force-pushing over a shared branch, stopping every process of one name on a machine, pushing a file of passwords and keys to a shared repository, sending an insulting email to all staff, and posting internal prices in public. Listing files, checking what has changed, running tests and reading a document all went straight through.
The comparison above uses the first generation, because it is the one we ran sample by sample against the general model. The classifier in tethercode today is a later one: an open model of about 4 billion parameters with a small trained part for each job. It reads every text twice, and if either reading gives even a one-in-five chance that the text is the most sensitive kind, it is treated as that.
| Full test, 8,148 samples | Today's classifier | The generation before it |
|---|---|---|
| Right answers | 98.2% | 98.0% |
| Most sensitive samples found | 99.86% | 99.73% |
The open model it is built on, with none of our training, scored 636 right out of the 1,000 comparison samples and found 188 of the 250 most sensitive. That is level with the general hosted model, which is the point: what makes the difference is the training, not the size.
The Kev comparison above is measured. So is one more result. The rest are being measured on real repositories, and we will publish the method and the raw numbers with them, including the ones that do not flatter us.
First result
On the same task, with tethercode doing the searching over its index instead of Claude running the searches itself. From our own tests; the tasks, the method and the raw numbers go here with the rest.
Being measured
The same tasks, run with the router and context from memory, and run without them.
Being measured
Edits where the code map flagged something the agent had not accounted for.
Being measured
Tokens an agent got from shared memory instead of having the project explained again.
Kev is one part of tethercode: who may see what, what may leave, and what may be done. Or leave your email and hear when it can be downloaded.