Stopping hidden instructions

Quarantine: instructions are not memories

Text that tries to give the agent orders is stored, locked away, and only released by a person.

What gets caught

Text that tries to override instructions, switch the agent's task, change its role, extract its hidden prompt, or force a specific answer. Before checking, AETHER-33 undoes common disguises: look-alike letters, invisible characters, odd spacing and simple encodings.

What quarantine means

• The memory is kept, not thrown away, so the attempt itself is on the record.
• Its status says quarantined and the reason says which pattern matched.
• Search never returns its text.
• Only someone with the reviewer secret can read it or release it, and they must give their name. The release is written into the log.

Why a secret and not just a name

If the agent could release its own quarantined memories, quarantine would be pointless. So the person who reviews sets a secret once and keeps it away from the agent. Set it before you give an agent access, because whoever sets it first holds it.

curl -X POST https://cainstudio.online/aether33/api/reviewer \
  -H "Content-Type: application/json" -d '{"secret":"choose-a-long-secret"}'

What it misses

This is pattern matching. A politely reworded attack, another language, or an attack that uses no obvious command words can get through. That is why the next rule matters most: even a bad memory that gets in can never grant permission. See “Memory is never permission.”