Skip to content

Security Behavior

AInternet security is layered. Each layer has a narrow job. Do not use one layer to replace another.

Layer Jobs

Layer Job
JIS prove actor identity now
TIBET record the causal action chain
Cortex decide whether an actor may perform an action
SNAFT enforce bilateral consent and semantic scope
MUX open or deny routes without leaking status
Airlock isolate risky work and return a verdict
Triage handle exceptional or risky actions without weakening guardrails
Trust Kernel runtime enforcement floor

SNAFT Is Not a Runtime Toggle

SNAFT is a guardrail. It should not be edited mid-flight to make an operation pass.

Bad pattern:

action blocked
operator weakens SNAFT rule live
action passes
audit trail claims policy was satisfied

Correct pattern:

action blocked
triage opens an explicit review path
operator approves, rejects or scopes exception
TIBET records the decision
SNAFT remains the guardrail

If a rule can only be installed by disabling the rule system, the system is not secure.

What Triage Does

Triage is the operator intervention layer. It handles work that is not safe to auto-allow.

Examples:

  • unknown actor requests a task
  • risky payload needs inspection
  • canary or fixture operation needs temporary opening
  • policy is incomplete
  • generated code wants execution
  • consent scope is ambiguous

Triage should create an explicit action trail:

incoming request
  -> risk classification
  -> hold/quarantine
  -> operator or policy review
  -> allow/deny/scope
  -> receipt

Airlock Behavior

Airlock is the isolation boundary for risky work.

Default behavior:

unknown -> quarantine
risky -> isolate
unsafe -> deny
approved -> execute with constraints and receipt

An Airlock run should record:

  • input hash
  • requested capability
  • policy decision
  • constraints
  • runtime posture
  • output hash
  • verdict
  • TIBET receipt

MUX and L0

MUX must avoid differential answers to unentitled callers.

world -> 0x0000
proven but unrelated -> 0x0000
proven and related -> scoped status or route

handshake.aint is the test fixture for reachability. Real actors should not be probed as test targets.

Fail-Closed Matrix

Condition Behavior
no actor proof deny or floor
stale proof deny or require freshness
unknown actor .paint / quarantine
no relation floor / no route
missing consent hold or deny
ambiguous intent triage
risky workload Airlock
failed Airlock verdict deny
missing receipt do not finalize action

Machine-Readable

  • https://ainternet.org/api.json — security gate, policy and route verbs
  • https://ainternet.org/resources.json — security behavior and conformance references