Context Poisoning
Context poisoning is when a wrong idea gets into the notes and goals an AI tool works from, and the tool keeps acting on it. Google's Gemini 2.5 technical report (2025) describes it in an AI agent playing Pokémon: false beliefs about the game spread into its goals and summary, and undoing them often took a long time.
Source: Gemini Team, Google, Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities (2025), Gemini Plays Pokémon section, checked October 29, 2026
What Context Poisoning means for your business
The report says the agent got stuck chasing goals it couldn't reach and repeated the same wrong statement, even when common sense said otherwise. Security writers also use the name for something else: attacks, where someone plants false facts or hidden orders for an AI tool to pick up.
The report's point is that once a wrong idea is in the tool's own notes, it keeps steering the task and is slow to undo. For a business, we'd read it this way: a wrong fact an AI tool picks up about you early in a task may be hard to shake later in that same task, so the first page it reads about you matters. That's our reading of the report, not something an AI company has said.
If an AI tool keeps repeating one wrong thing about you, finding the page it keeps leaning on is our first step.
Last reviewed October 29, 2026. All terms
How AI answers it today
On October 29, 2026, Google's AI Overview cited data and security companies' blogs and IBM's docs, and covered both mistakes and attacks. ChatGPT treated it mostly as an attack where someone plants bad information, and cited nothing.
Our run through ChatGPT and Google's AI Overview, October 29, 2026. A sample, not what every buyer sees.
Related terms in answers and how content gets used
Find out where your brand stands.
Tell us what you're trying to figure out. We'll get back to you within one business day with a time to talk.