Black Friday PreviewAmazon USBlack Friday preview: compare before the rush hitsAmazon US: current deals, useful picks and tech finds.Check DealsSlow PC?RecommendedPC slow today? Run a repair scan before it gets worseResolve common Windows issues and optimize system performance.Scan NowBlack Friday PreviewAmazon USThe best pick may show up before FridayAmazon US: current deals, useful picks and tech finds.See Picks×
FREEDOM251

Freedom251 › Blog

AI Worms Explained: How Malicious Instructions Jump Between Agents—and How to Block Them

By Freedom251 Team · · 10 min read

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.

An AI worm is a security attack in which malicious instructions hidden in content read by one AI agent can influence that agent and then travel through its normal messages, retrieved data, shared memory or tool outputs to another agent. The “worm” analogy describes this instruction-level propagation, not a conventional executable virus. A spread requires a vulnerable agent, a path into its context and enough permissions or connectivity to relay the attacker’s content or actions. Research has demonstrated the pattern in controlled retrieval-augmented generation (RAG) and email-assistant scenarios, but there is no reliable figure showing that AI-worm outbreaks are common in production.

What is an AI worm?

An AI worm is a worm-like attack against systems that use large language models as agents. The attacker plants instructions inside data the agent is expected to read—such as an email, document, web page, calendar entry or tool result. If the model treats those instructions as part of its task rather than as untrusted data, it may perform an unintended action or include the instructions in content sent to another agent.

As an Amazon Associate I earn from qualifying purchases.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

That differs from a traditional computer worm. A conventional worm usually exploits software vulnerabilities and copies executable code between machines. An AI worm can move as text or other model-readable content through legitimate application paths. The receiving agent does not need to download a binary or run malware; it only needs to ingest attacker-controlled content and give it enough influence over its reasoning or tool choices.

The term describes a mechanism and a risk pattern, not proof of a widespread incident class. The 2024 Morris II research paper demonstrated worm-like propagation in a RAG-based email-assistant ecosystem. Its results apply to the described experimental setup, not automatically to every deployed agent.

Popular on Amazon
Silkland 40Gbps USB C 90 Degree Right Angle Adapter 2 Pack, 240W USB4 | Male to Female Extender, 8K@60Hz, Thunderbolt 4/3, Steam Deck, ROG Ally, MacBook, Switch 2, iPhone 18/17/16/15, Quest 3 VR
  • 【90 Degree Full Function USB C Adapter】Supports the latest USB4 standard for up to 40Gbps of bandwidth, 8K video support, and up to 240W Fast charging; Support USB PD 3.1 Extended Power Range (EPR) including 140W, 180W and 240W. Your productivity is easily doubled. *Warm Tip: For the best fit, please make sure the USB-C port is fully exposed; thick or bulky cases may affect compatibility.
  • 【Reliable Quality】Made from high-quality, scratch-resistant aluminum alloy, delicate metal frame. get excellent wear resistance and heat dissipation. Non-Slip Design for easier insertion and removal.
  • 【Small and portable】Ultra thin size, ultra compact design, a petite physique that matches the minimalist design of your modern laptops. Coming with a slim figure that won't block the port next to the one it's plugged into.
  • 【Compatibility】Widely work with most usb c, thunderbolt 4/3 devices. Compatibilit for ROG Ally, Steam Deck, Switch, MacBook Pro, Quest 3 VR headset, Surface Book 2, iPhone 18/17/16/15, Samsung Galaxy S23/S22/S21, Note 20/10, Google Pixel, ChromeBook, etc.
  • 【Excellence Customer Support】If you experience any issues or cable failure, simply reach out to our dedicated customer service team. We provide a prompt response within 18 hours, providing you with the necessary technical support to resolve any concerns.

How does an AI worm spread between agents?

Propagation is conditional. An agent must read the malicious content, the content must affect its behavior, and the system must provide a route for the resulting text or action to reach another workflow.

  1. Content is seeded. An attacker places an instruction in an email, shared document, web page, support ticket, database record or another source likely to be retrieved. The payload may look like ordinary prose and does not have to be executable code.
  2. An agent ingests it. A retrieval system, connector or tool fetches the content and places it in the model’s context. In a RAG application, the retrieved passage can appear beside system instructions and the user’s request.
  3. The model is hijacked. The embedded text tells the agent to ignore its original task, reveal information, alter a record, call a tool or reproduce the instruction elsewhere. NIST calls this form of indirect prompt injection agent hijacking.
  4. The agent takes or generates an action. If permitted, it might send a message, create a file, update a database, invoke another tool or summarize the malicious text. The action can be harmful even if no other agent is involved.
  5. The content crosses a boundary. Messages, generated summaries, shared memory, retrieved records and agent-to-agent handoffs can carry the attacker’s wording or an equivalent instruction into another agent’s context.
  6. A downstream agent is exposed. If that agent also fails to distinguish external data from control instructions, it may repeat the cycle. Each hop depends on the next agent’s connectors, permissions and safeguards.

This is why “spread” should not be read as automatic or instantaneous. Remove the retrieval path, block outbound messaging, require approval for the action or sanitize the content, and the chain can stop. Conversely, broad permissions and many connected agents increase the number of possible routes and the consequences of a single compromise.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

What is indirect prompt injection, and why is it central to AI worms?

Indirect prompt injection occurs when an attacker puts instructions in data an AI system consumes indirectly, rather than typing the attack into the user’s chat box. The application retrieves or displays the data, and the model then interprets the attacker’s text as an instruction. In its January 2025 technical blog on agent-hijacking evaluations, NIST states that many agents are vulnerable to this pattern.

The underlying trust-boundary mistake is treating all text in the model’s context as equally authoritative. System and developer instructions are intended to control behavior; an email or web page should normally be treated as evidence or content to analyze. Language models, however, receive both as tokens in one context window. If the application does not enforce the distinction in its architecture and tool layer, persuasive text can influence the model’s next action.

Also popular on Amazon
Sale
Anker USB C Hub, 5-in-1 USBC to HDMI Splitter with 4K Display | 85W Max PD, 1 Powered USB-C 5Gbps & 2 Powered USB-A 3.0 5Gbps Data Ports, for MacBook Pro, MacBook Air, Dell and More (Black)
  • 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
  • Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
  • Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
  • HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
  • What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.

An AI worm adds a propagation objective to that injection. The payload might tell an email agent to forward a similar message, instruct a document-processing agent to insert text into a shared file, or cause a coordinator agent to pass a crafted summary to a specialist. The exact wording is less important than the existence of a relay path.

Does an AI worm require someone to click?

Not necessarily. A user may never click a link if an automated agent polls a mailbox, indexes a drive, fetches a web page or processes a support ticket on its own. That is why the Morris II paper used the phrase “zero-click” for a scenario in which normal automated retrieval supplied the malicious content.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

“Zero-click” does not mean zero conditions. The attacker still needs a way to place content where the agent will read it, and the target system must automatically process that source. If a workflow requires a human to approve every message, blocks external senders or never forwards model-generated text, the same payload may fail to propagate. Automation changes the trigger from a person’s click to a scheduled or event-driven data flow.

Which channels can carry an AI worm?

Any channel that moves model-readable content across a trust boundary can become a carrier. Common examples include:

Also popular on Amazon
Sale
Anker USB C Hub, 7in1 Multi-Port USB Adapter, 4K@60Hz USBC to HDMI Splitter | 85W Max PD, 2 USB 3.0 & 1 USBC Data Ports, SD/TF Card Reader, for Laptops/Mac and Other Type C Devices (Charger Not Included)
  • Sleek 7-in-1 USB-C Hub: Features an HDMI port, two USB-A 3.0 ports, and a USB-C data port, each providing 5Gbps transfer speeds. It also includes a USB-C PD input port for charging up to 100W and dual SD and TF card slots, all in a compact design.
  • Flawless 4K@60Hz Video with HDMI: Delivers exceptional clarity and smoothness with its 4K@60Hz HDMI port, making it ideal for high-definition presentations and entertainment. (Note: Only the HDMI port supports video projection; the USB-C port is for data transfer only.)
  • Double Up on Efficiency: The two USB-A 3.0 ports and a USB-C port support a fast 5Gbps data rate, significantly boosting your transfer speeds and improving productivity.
  • Fast and Reliable 85W Charging: Offers high-capacity, speedy charging for laptops up to 85W, so you spend less time tethered to an outlet and more time being productive.
  • What You Get: Anker USB-C Hub (7-in-1), welcome guide, 18-month warranty, and our friendly customer service.
  • Email and messaging: an agent may read an incoming message and send a reply or notification to another agent or contact.
  • RAG indexes: a poisoned document can be embedded, retrieved later and inserted into many otherwise unrelated tasks.
  • Shared files and databases: generated notes, tickets or records can be consumed by downstream workflows.
  • Tool results: browser, search, CRM or code-analysis tools may return attacker-controlled text that the model trusts.
  • Agent handoffs: a coordinator can pass a specialist’s summary, plan or recommendation to another model without preserving the original trust label.
  • Shared memory: persistent conversation or task memory can preserve malicious instructions beyond the first interaction.

The carrier may be transformed at each hop. A downstream model does not need to receive the exact original string; a summary that preserves the attacker’s instruction can be enough. Conversely, retrieval filters, content labeling, output validation or a human review step may break the chain.

What did the Morris II research actually show?

Stav Cohen, Ron Bitton and Ben Nassi’s 2024 “Here Comes The AI Worm” paper studied self-replicating adversarial prompts in a GenAI email-assistant ecosystem using RAG. The work examined factors such as context size, embedding configuration and propagation hops, and investigated confidential-data extraction.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Those are experimental findings from a specified architecture. They show that a malicious prompt can be designed to replicate through connected AI applications under suitable conditions. They do not establish how often this occurs in deployed products, that every agent is vulnerable, or that a successful first hop guarantees an outbreak. The available data contains no reliable production-prevalence statistic.

How difficult is it to defend tool-using agents?

Security evaluation remains difficult because an agent must satisfy its normal task while resisting adversarial content. The NeurIPS 2024 AgentDojo benchmark contains 97 realistic tasks and 629 security test cases for tool-using agents. Its results show that prompt-injection attacks can break some security properties and that both attacks and defenses remain challenging.

Also popular on Amazon
BENFEI USB C Hub 5-in-1 with 4K HDMI(Certified), 100W Power Delivery, 3 USB-A, Silicone Cable, Aluminum Case Compatible with MacBook Pro/Air, iPad Pro, iMac, iPhone 15 Pro/Pro Max, XPS, Thinkpad
  • Portable and powerful USB-C HUB: BENFEI USB Type-C HUB, with super-soft and knot-free silicone woven design cable, meets most mobile office needs. Compact, lightweight, stylish, and powerful portable USB C Hub equipped with 1 x HDMI port, 1 x 100W charging, and 3 x USB ports. 18-month warranty, 24-hour response, to ensure you feel at ease when using our product.
  • Design centered on comfort and reliability: Thanks to BENFEI's end-to-end in-house cable production capability, in-house PCBA and assembly capability, using the industry's most advanced silicone woven design and process, 20cm cable in length, no knots, super-soft, the HUB is easy to use in all scenarios: laptop, tablet, stand etc. Super-soft, 25000+ life cycles, to meet your daily carrying and office needs.
  • 100W Charging: Support up to 90W USB C pass-through charging via Type-C port to keep your laptop powered. 10W is reserved for other interface operations. No data and video function on the Type-C port.
  • 4K HDMI Display: The HDMI port supports media display at resolutions up to 4K 30Hz, keeping every incredible moment detailed and ultra vivid. Please note that the C port of the Host device needs to support video output.
  • Transfer Files in Seconds: Transfer files and from your laptop at speeds up to 10 Gbps with USB A 3.2 port. Extra 2 USB A 2.0 ports are perfectly for your keyboards and mouse.

The benchmark counts are not incident statistics, affected-user counts or a certification. A system that performs well on a benchmark can still fail when its own connectors, data and permissions create different attack paths. Evaluation should therefore resemble the production workflow: use the real tools, data formats, approval rules and agent handoffs, then test malicious content in each external source.

Could newer AI research make worms more adaptive?

A June 2026 arXiv preprint titled “AI Agents Enable Adaptive Computer Worms” studies agents that generate tailored attack strategies. It is preliminary research, not evidence that adaptive AI worms are prevalent in real-world systems. Its relevance is strategic: agents may eventually help attackers adjust payloads to different software, contexts or defenses, increasing the value of strong isolation and monitoring.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

How can an organization reduce AI-worm risk?

No single filter guarantees that prompt injection is eliminated. Effective reduction combines least privilege, explicit trust boundaries, approval controls, isolation and realistic evaluation.

  1. Inventory every untrusted input. List the mailboxes, websites, files, tickets, search results and tool responses each agent can read. Mark which sources are attacker-controllable and how often they are fetched.
  2. Separate data from control instructions. Keep system policies outside retrieved text, label external content as untrusted and design prompts and tool wrappers so a document cannot redefine the agent’s role. Treat labels as a defense layer, not as proof that the model will obey them.
  3. Apply least privilege. Give an agent only the tools and data required for its task. Read-only access is safer than write access; a draft-message capability is safer than unrestricted sending; a scoped record update is safer than database-wide administration.
  4. Require approval for consequential actions. Human or policy approval should gate external messages, financial changes, account permissions, deletion, code deployment and bulk edits. Show the source content and proposed action to the reviewer.
  5. Constrain cross-agent communication. Use allowlisted recipients, typed message schemas, provenance labels and size or rate limits. Do not let one agent silently promote another agent’s output to trusted instructions.
  6. Monitor for propagation signals. Alert on unexpected outbound messages, repeated instruction-like text, unusual tool sequences, access to unrelated records, rapid fan-out and attempts to disable safeguards. Preserve logs linking each action to its input source.
  7. Evaluate realistic attacks. Test malicious emails, poisoned documents, hostile web results and crafted tool responses in workflows modeled on actual deployments. Include multi-hop handoffs and recovery after a suspected compromise; AgentDojo can inform methodology but is not a universal safety seal.
  8. Limit blast radius. Isolate tenants and workloads, use short-lived credentials, rotate tokens after suspected exposure and make memory and indexes easy to purge or restore. A compromised helper should not automatically control the coordinator or every connected system.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Support on Ko-Fi

What should you do if an agent may be hijacked?

First pause the suspected workflow’s outbound actions and revoke or narrow its tool credentials. Preserve prompts, retrieved documents, tool-call logs, message recipients and timestamps before deleting poisoned content. Then identify every downstream agent, index, file or recipient that consumed the output.

Also popular on Amazon
Sale
𝟒𝐏𝐚𝐜𝐤 𝐔𝐩𝐠𝐫𝐚𝐝𝐞𝐝 USB to USB C Adapter,Type-C Female to A Male Charger Converter for iPhone 18 Charger,iWatch Ultra SE 9 10,iPhone 17 Pro Max 16 15 14,Samsung S25,AirPods,iPad,Car,Xbox
  • 【Read before You Buy】 Note: This adapter is USB 2.0 version and cannot transmit HDM, VGA, or other video signals. Nevertheless, they are fully capable of charging devices,duplicating files, and can be seamlessly integrated with USB-C earphones, wireless network adapters, and SD card adapters.
  • 【Easy to Use and Safe】Thanks to modern manufacturing techniques, we've made these adapters sufficiently compact so that they can remain inserted in your device without noticeable protrusion or inconvenience. This means you don't have to constantly plug and unplug them, reducing the likelihood of wear and misalignment, and making them more convenient to carry around.
  • 【Reliable Quality】Engineered for enhanced durability, durable aluminum alloy shell material,get excellent wear resistance and heat dissipation,the male USB C and female USB connectors & the male USB and female USB C connectors are designed to be plugged and unplugged up to 20000 times. Provides data transfer, device charging, memory space expansion, and additional computer peripherals.
  • 【Think more for you】Do you need to connect a USB-C-enabled device to an odd device that only has USB-A ports? Then these USB-A male to USB-C female adapters are made for you.Simply insert the adapter into the USB-A port on the device, and you can instantly connect your USB-C device to the device that originally didn't have the correct connection port.
  • 【Trusted xiwxi Quality and Service】As a responsible brand seller, you will receive xiwxi's 12-month Quality assurance and lifetime friendly customer service, you're not just buying a product,you're investing in peace of mind with all-round quality and exceptional customer service.

Remove or quarantine the malicious source, invalidate contaminated memory or retrieval records, and review actions taken under the agent’s credentials. Rotate secrets if confidential data or privileged tools may have been exposed. Re-enable automation gradually, beginning with read-only operation and human approval, while replaying the incident as a regression test.

What an AI worm is not

  • It is not automatically a conventional virus or executable malware.
  • It is not proof that every language-model agent can infect every other agent.
  • It is not necessarily dependent on a user clicking a link.
  • It is not measured by the 97-task, 629-test AgentDojo benchmark; those are evaluation-set sizes.
  • It is not a claim that production outbreaks are frequent; current evidence is primarily controlled research and benchmarks.

The practical conclusion is narrower and more useful: whenever an agent consumes untrusted content and can pass model-generated text or actions into another workflow, prompt injection can become a propagation problem. Designing explicit trust boundaries and limiting autonomy reduces both the chance of a successful hop and the damage if one occurs.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Frequently Asked Questions

Can a prompt injection spread from one AI agent to another?

Yes, when the first agent’s output is delivered to a second agent as readable context and the second agent treats embedded instructions as authoritative. The transfer can occur through a message, summary, shared memory, retrieved record or tool result; it is not guaranteed when those channels are filtered or approval-gated.

What is the difference between an AI worm and ordinary prompt injection?

Ordinary prompt injection targets one model’s behavior. An AI worm adds a relay or replication step, using the compromised agent’s normal communication or data pathways to expose other agents. A single injected task can therefore be harmful without being worm-like if no onward path exists.

Can retrieval-augmented generation prevent AI worms?

RAG is not a security boundary by itself. Retrieval can supply useful evidence, but it can also import attacker-controlled instructions; isolation, provenance, least-privilege tools and approval checks are needed around the retrieved text.

How can I tell whether an agent passed on malicious instructions?

Review message and tool-call logs for unexpected recipients, repeated instruction-like phrases, unusual fan-out, access outside the task scope and actions immediately following a new external document or tool result. Trace the provenance of each downstream input instead of looking only at the final response.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Are AI worms already common in production?

The cited research does not provide a reliable production-prevalence number. Morris II is a controlled demonstration, AgentDojo is a benchmark, and the adaptive-worm work is a preliminary preprint, so they support the risk mechanism rather than a claim of widespread outbreaks.

Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply. Prices as of November 16, 2026 4:56 am GMT+0530.

Leave a Reply

Your email address will not be published. Required fields are marked *

Free tools Windows power users keep installed

One-click scans. No signup required.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Recommended PC Tool
Recommended PC Tool
Windows Errors? Fix Them Before They SpreadFree repair scan
Outdated Drivers Are Slowing You DownFree scan - exact matches

Two free Windows tools

One Free Minute Could Fix That PC

Before you go - each of these free tools takes about a minute and tackles what quietly slows a Windows PC down.

Special offer. View Outbyte info, uninstall instructions, EULA, and Privacy Policy.