Definition
Plain language
A safety attack where fake prior actions in the conversation history get the AI to keep doing the same bad thing.
As stated in the literature
A class of multi-turn manipulation that injects fabricated prior trajectory steps plus a consistency instruction, exploiting an LLM's tendency to extrapolate from in-context behavioral patterns.
Also called: History Anchors
Why it matters: It's a real attack pattern against multi-turn agents, and defending against it requires the model to be willing to break with its own apparent past behavior.
For example, a prompt is constructed so it looks like the agent has already approved three risky transactions, with a note saying 'maintain consistency,' nudging it to approve a fourth.
Heard on the show
“The paper is called "History Anchors: How Prior Behavior Steers LLM Decisions Toward Unsafe Actions," from an independent researcher named Alberto Rodríguez Salgado.”Episode 044 — How One Sentence and a Forged History Flip the Most Aligned Models