Transom

Context poisoning

Also called: context drift, memory poisoning

Corrupting the model's working context — its conversation history, retrieved documents, or persistent memory — so later steps inherit attacker-controlled content.

An agent accumulates context: past turns, tool outputs, notes it saved to memory, documents it retrieved. If any of that is attacker-influenced, the poison persists and compounds — a malicious instruction written into long-term memory on turn 3 keeps acting on turn 30. "Context drift" is the softer version, where accumulated noise and stale content degrade the agent's grounding over a long session.

Why it matters

It turns a one-shot injection into a persistent foothold and makes incidents hard to diagnose, because the triggering content is no longer on screen. Memory writes deserve the same scrutiny as tool calls, and long-running agents benefit from context that can be audited and reset.

Exposure map

Live counts of instances showing this pattern will appear here once the exposure map is collecting data.

Related terms

Sources

Added 2026-09-02. Last reviewed 2026-09-08. Definitions in this space are evolving; entries are dated so revisions stay legible.