Agent data exfiltration

An agent moves protected information to an unauthorized recipient or destination.

01 / Attack path

How the boundary is crossed

A finance agent reads a ledger for reconciliation, then sends it to an external mailbox supplied by a document comment.

Attack path and interception point
  1. 01Lower-trust input
  2. 02Attempted instruction
  3. 03Protected action
  4. 04Enforced decision

02 / Containment

Where bounded authority limits the effect

Authorize export separately from read, constrain recipient and data classification, and block the send at the mail or network executor.

03 / Failure and evidence

The attacker’s opportunity and the defender’s record

Exploit condition

The read credential and send credential together create an unreviewed data-transfer capability.

Evidence to retain

Correlate source record, destination, content digest, decision and outbound delivery result without logging the ledger.

04 / Canonical scope

Why this reference stands alone

Prompt injection can induce the attempt; this page owns the protected-data egress effect.

Protocol or attack trace

Read-to-send escalation

Sequence

Agent reads ledger under reconciliation task; document comment supplies external mailbox as export recipient.

Negative test

Mail executor requires distinct export grant for ledger class and exact recipient; deny mere read authority.

Primary references

Read the underlying material

Architecture discussion

Choose one consequential action and make its boundary explicit.

Request a Conversation