AI ACTION SAFETY LAYERFAIL CLOSED BY DEFAULT

Verify the context.
Prove the lineage.
Then let the agent act.

Quarantine uses HydraDB provenance graphs to determine whether an AI-generated action is backed by trustworthy, resolvable evidence. If the lineage cannot be proven, the action is blocked.

01
EvidenceUntrusted input
02
ProvenanceHydraDB lineage
03
VerificationSigned graph state
04
PolicyDeterministic rules
05
Action gateAllow or block

Two inputs. Same pipeline. Opposite outcomes.

Run the trusted flow, then attack the same boundary with tampered provenance.

ACTION GATE RESULT

Verifying live provenance

The adapter cannot run until HydraDB provenance, trusted state, policy, freshness, and replay checks pass.

Provenance
VERIFYING
Policy
WAITING
Gateway
WAITING
Adapter invocations
0
Action
-
Destination
-
Proof decision
-
Independent paths
-
Trusted sources
-
Authorization
-
01

INCOMING EVIDENCE

What the agent wants to act on

UNTRUSTED
CLAIM

Waiting for the trusted demo runner…

Source
-
Entity
-
Received
-
Input state
-
AI boundary

The producer supplies content only. It cannot submit trust, verification, policy, ancestry, or an adapter.

02

HYDRADB PROVENANCE

Trace the evidence to its sources

WAITING
Nodes
-
Edges
-
Witnesses
-
Paths
-
Observed depth
-
Auth bound
-
Live HydraDB provenance graph The graph is populated from the trusted server response.

Querying the live HydraDB provenance closure.

trusted source derived evidence action argument displayed source → action

Displayed source-to-action for readability; HydraDB stores child-to-parent DERIVES_FROM edges.

No provenance witnesses loaded.
03

DETERMINISTIC DECISION

The model does not authorize this

WAITING
FINAL SECURITY BOUNDARY

Action gateway

NOT RUN
Awaiting graph verificationAction adapter has not been called.

Provenance verification

NOT RUN
  • Waiting for live graph state-

The trusted verifier waits for a graph-backed result.

Deterministic policy

NOT RUN
  • Verified trusted state
  • Allowed source authority
  • Internal destination
  • Fresh request
  • No replay

Policy is evaluated only after provenance verification succeeds.

SERVER EXECUTION TRACE

What actually happened

No run yet
  1. The trusted execution trace will appear after a run.

Similarity is not provenance.

Vector search can find information that looks relevant. Quarantine needs a stronger answer: can this proposed action be traced through a valid chain of authenticated evidence?

RETRIEVAL

“This text looks related.”

Not enough to act
HYDRADB PROVENANCE

“These signed relationships form a resolved lineage to trusted sources.”

Can enter policy
HydraDB is not a storage plugin. It is part of the authorization path.