What's this about?
The Guardian reports on a fabricated American thinktank, set up and funded by Israel, whose purpose included seeding material intended to be picked up by AI systems as authoritative source material.
What exactly is the attack here?
It is not prompt injection — it is upstream of that. Models and their retrieval systems treat institutional sources as credible. If you manufacture an institution, publish through it, and get indexed, you are not tricking the model at query time; you have poisoned the pool it draws from. It works against search engines and journalists by the same mechanism, and AI simply industrialises the payoff.
What's the community saying?
The discussion focuses on how ordinary the technique is — fake thinktanks predate AI entirely — and on the newly high return on investment: one manufactured source, once indexed, can be repeated by every model that ingests it, at no marginal cost to the operator.