The terminology depends on the speaker. Michel Tricot uses it on the show for a reliability problem: an agent receives too much unhelpful data and struggles with the task. Security writers may use the same term for deliberate manipulation. Name the mechanism rather than assume every instance is an attack.
In episode 52, at 9:02, Tricot discusses post-processing API data before it reaches an agent. Conor’s episode write-up compares a targeted retrieval path with raw paginated output.
Example: before upgrading the model, inspect whether a failed run received the right records, the right version, and enough evidence to answer. Then compare a smaller relevant context against the original on the same task. This is a diagnostic experiment, not proof that the model is never the problem.
Long-context research also shows that information position can affect retrieval performance. That supports testing context use, but does not establish a single cause for every failure. Prompt injection covers instructions planted in untrusted input.
Hear it from the guest
“A lot of this is done on the back end to make sure that we only return what we believe is the most relevant information. So that's the first part. And that's one thing you will have a hard time doing with APIs, because it will just return to you the whole block of information.”
Quotes lightly edited to remove filler words.
Sources
- Liu et al.: Lost in the Middle — Shows position-dependent performance in long-context tasks; it does not define every use of context poisoning.
- Conor Bronsdon: Your AI Agents Are Drowning in Bad Context — Documents the show's reliability-focused use of the term and the API-output example.
