Agent memory contamination

Untrusted content stored as long-term memory can influence later unrelated tasks.

01 / Attack path

How the boundary is crossed

A support ticket embeds a false instruction that the agent summarizes into its persistent customer-memory store.

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

Separate observations from instructions, attach provenance and expiry to memory, review sensitive writes and recheck every future action against the current task.

03 / Failure and evidence

The attacker’s opportunity and the defender’s record

Exploit condition

A stale attacker-authored memory item is treated as an enduring business rule.

Evidence to retain

Retain memory source, write approval, later retrieval reference and blocked action.

04 / Canonical scope

Why this reference stands alone

Tool-output laundering affects an immediate result; this page owns persistence across sessions.

Protocol or attack trace

Persistent instruction seed

Sequence

Ticket embeds false policy; summarizer writes it to long-term memory and later task retrieves it.

Negative test

Memory write retains provenance and cannot create authorization; later action still requires current trusted task.

Primary references

Read the underlying material

Architecture discussion

Choose one consequential action and make its boundary explicit.

Request a Conversation