Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJan 11, 2026Singapore
4th place

Agent Attacks via Memory Injection

Leonidas Raghav, Choong Kai Zhe · Team Mem::Poison

Submitted to AI Manipulation Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Agent Attacks via Memory Injection

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

LLMs have recently adopted persistent memory to provide models with better user knowledge and personalisation. However, this introduces a new vector for adversarial manipulation. This report investigates Memory Injection, a threat model where adversaries exploit indirect prompt injection within web content to poison an agent's long-term memory. Employing user manipulation scenarios, we show that memory attacks, should they succeed, are effective at changing model behaviour towards the user, often more so than a direct system prompt. Hence, this highlights the need for more robust evaluations of memory updates in agentic memory systems.

This work was inspired by the SPAR Spring 2026 project 'Latent (Sleeper) Attacks via Persistent Memory', proposed by Ivaxi Sheth (CISPA Helmholtz Center for Information Security) and Vyas Raina (University of Cambridge). We are grateful to Ivaxi and Vyas for this research direction, and to SPAR (Supervised Program for Alignment Research) for enabling its publication.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. This seems like good quality and important AI security research! Honestly I couldn’t quite tell if it aligns with the “manipulation” scope of the hackathon (because it feels like an even more general and important security vulnerability), but in any case the authors do focus on the domain of manipulation, and the attack vector (poisoning persistent memory via hidden web content) is for sure very concerning for agentic systems. The findings are honestly pretty crazy (100% ASR on memory injection attacks for grok-4-fast??). It’s good at least to see that GPT-4.1 is doing better 😬

  2. This is an interesting and underexplored area: how prompt injection could be used to manipulate users. The design seems appropriate and the results were very striking! The authors did a great job at testing follow up explanations / hypotheses.

    I think the submission could be strengthened by providing more details about the mechanics of the injection (in particular whether the agent scaffold used native APIs which I would expect to have stronger guardrails) or external. In addition an appendix with more examples of the exact prompts and web text would be useful.

    In future work it would be great to see the authors measure the effectiveness of the manipulation on real users though the LLM-as-judge is a great start.

Cite this project

@misc{raghav2026agent,
  title = {{Agent Attacks via Memory Injection}},
  author = {Leonidas Raghav and Choong Kai Zhe},
  year = {2026},
  month = jan,
  note = {Submitted to AI Manipulation Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/agent-attacks-via-memory-injection-izfz}},
  url = {https://apartresearch.com/sprints/projects/agent-attacks-via-memory-injection-izfz}
}

More from AI Manipulation Hackathon

All 70 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026