Part 4 of 13 · 2 min read
Break the trifecta: no one agent should hold all three dangerous properties
An agent that can read anything, see sensitive data, and act on the outside world is a loaded gun pointed at whoever controls the text it reads.
Think of a bank teller who reads anonymous notes from the public, has access to every customer's account balance, and can wire money anywhere with no second signature. The danger is not any one of those three abilities. The danger is holding all three at once.
the lethal trifecta is the combination of private data, untrusted content and external communication in one agent, and the rule of two is the concrete policy built from it. Allow at most two of untrusted input, sensitive access, or state change/external communication per session -- which breaks the dangerous combination by construction, in code, rather than hoping the model behaves.
A worked example. In the example's simulation, with no policy at all, an injected instruction leaks the private customer table 28% of the time at an assumed 30% model-obedience rate, and 92% of the time at 100% obedience. With the rule-of-two policy enforced, the leak rate is exactly 0% at both obedience levels -- because the policy denies the send_email call the moment it would make the session satisfy all three properties (A: it already read untrusted web text, B: it already read the private customer table, so C: sending email anywhere but the owner's address is refused), regardless of whether the model "wanted" to send it.
The key lines. From code/mini-injection/injection.py:
# from code/mini-injection/injection.py
def rule_of_two(session, tool):
"""Deny any call that would make the session satisfy all three properties (no human approval here)."""
a, b, c = TOOLS[tool]
f = {k: session.flags[k] for k in "ABC"}
f["A"] |= a; f["B"] |= b; f["C"] |= c
return not (f["A"] and f["B"] and f["C"])
Read it as: the function checks what the session's flags would become if this call were allowed, before allowing it. It is not asking "is this specific call suspicious?" It is asking "would allowing this call complete the dangerous triple?" That is why it works identically whether the model is innocently confused or maximally obedient to an attacker. The check never looks at intent. It looks only at the structural combination of powers.
Lethal trifecta, rule of two
The lethal trifecta is private data, untrusted content and external communication in one agent. The rule of two is the policy of allowing at most two of untrusted input, sensitive access, and state change or external communication per session.
Example: the leak rate collapsing from 92% to 0% under 100%-obedience conditions is the rule of two doing exactly what a filter cannot: it doesn't try to detect the attack, it makes the attack's final step structurally impossible regardless of whether the model obeyed every single injected word along the way.