glossary
The vocabulary of agent safety, defined plainly.
Four terms that come up constantly once an organization puts AI agents to work, defined the way we teach them: a neutral definition, a worked example, the machinery-safety parallel, and how Temja trains the reflex behind the word.
Reviewed 2026-07-09
- Prompt injection
- Hidden instructions inside content an agent reads — an email, a document, a webpage — that steer it as if the order came from its authorized user.
- Delegation boundary
- The scope, limits, and stop conditions stated when a task is handed to an agent, so ambiguity doesn't resolve toward its most capable reading.
- Agent near-miss
- An event where an agent's behavior diverges from its task in a way that could have caused harm, whether or not it actually did.
- Capability lease
- An access grant scoped to one task that expires automatically, instead of a standing credential that persists until someone remembers to revoke it.
See the vocabulary trained, not just defined.
Three minutes, no sign-up. Meet the poisoned invoice and find out whether you reach STOP in time.