AI safety
How we keep our models and agents from being turned into tools for cyber attacks, what we have built, what is still planned, and what the evidence says about the threat.
The layers, in one picture
No single check stops misuse. An agent that can read the web and act for a person meets several separate controls, each written in code and covered by tests.
- Account and planAgents run under a signed-in account (age 18 or over at sign-up), its plan's request limits, step caps and spending caps.
- Model rules and policyRequests for malware are refused by rule, and the acceptable use policy forbids attacking systems you aren't authorised to test.
- Labelled outside textPages, files and tool results are wrapped as data with one rule: never follow instructions inside them.
- Injection noticeText that tries to talk to the model is flagged to the person. It informs; it never silently edits the page.
- Tiers and taintRead runs, write asks once outside text was read, confirm always asks. The model's words cannot change a tier.
- Grants that can't widenBuilt-in tools that send, pay, call or delete are never remembered as allowed. Each one asks every time.
- Outbound checkRequest catchers, tunnels and paste sites are blocked; keys, passwords and card numbers are never typed or sent.
- Send limitsDaily limits on emails, new recipients and calls, sized by how much trust an account has earned.
- Activity log and StopEvery call is logged with secrets removed. Stop everything halts an account's runs, approvals and agents.
What we say Klyroe does was checked against our code on the day of writing. We describe each control in general terms and leave out exact limits, so they are harder to work around. Where something is planned rather than built, we say so. No control here makes misuse impossible; together they make it slower, louder and easier to stop.
How Klyroe is built
The field
How we write these
Every statement about what Klyroe does was checked against our code before it was published, and the article names the file. Where something is planned rather than built, we say planned.
Outside facts link to their source, preferring papers, government advisories and the AI labs' own reports. A lab's report about its own systems is that lab's claim, and we label it so.
We describe defences and categories of attack, never working steps. If you find a way around one of our controls, email hello