An “AI worm” can spread through Microsoft Word documents using Copilot as a vector, a prominent Norwegian AI researcher reported on Tuesday.
The report from Håkon Måløy, later confirmed by Microsoft, said that an attacker can conceal instructions in a document that is later used as source material for Copilot-generated or Copilot-edited Word documents, for example, as input to a financial report. Those malicious instructions could potentially alter figures in the document being created. They then copy themselves into the new document, causing it to become a carrier of the attack when used in another Copilot-assisted workflow.
Måløy noted: “To my knowledge, this is among the first public demonstrations of document-borne AI-worm self-propagation through normal workflows in a mainstream commercial productivity suite.”
Microsoft on Thursday emailed a statement to CSOonline discussing the report’s revelations.
“We have addressed the findings reported by the researcher and thank them for working with us through coordinated vulnerability disclosure. To address this class of risk, we use a defense-in-depth strategy with safeguards that block malicious instructions at multiple points and help keep tasks aligned with users’ requests,” Microsoft said.
“We are continuously strengthening these safeguards as the technology and threat landscape evolve,” it added. “We encourage customers to install the latest updates, use multiple layers of security protection, treat content from unknown sources with caution, and review AI-generated content before using or sharing it.”
Sidesteps defenses
Aman Mahapatra, chief strategy officer for Tribeca Softtech, a New York City-based technology consulting firm, has reviewed the mechanisms used in this vulnerability and says that it is much worse than it might seem, because it will sidestep just about every defensive mechanism enterprises have in place today.
“This is a worm, a self-propagating malware pattern that uses Copilot as the transmission mechanism and legitimate corporate collaboration as the delivery channel. It bypasses every traditional email security control because the document is not malicious on delivery and it becomes malicious when Copilot processes it. That means that it bypasses DLP because the exfiltration happens through the user’s own authenticated Copilot session,” Mahapatra said. “It then bypasses endpoint protection because no code executes, only instructions get followed by an AI service the enterprise has explicitly authorized.”
He also noted that researchers have been warning about this class of attack for two years.
Microsoft’s involvement
Måløy said that he has been working with the Microsoft Security Response Center (MSRC) since March 3 and Microsoft subsequently implemented and distributed multiple small focused mitigations, but the core vulnerability has yet to be fixed.
He said he was hesitant to disclose an active vulnerability, but felt that it was now time to publish.
“My reasoning is that defenders cannot reduce exposure to a risk they are unaware of,” Måløy wrote, “and the propagation mechanism described here affects ordinary document workflows that many organizations already rely on.”
In an email interview, Måløy noted that the tweaks Microsoft has implemented were helpful, nonetheless.
“The coordinated disclosure process showed that mitigations can meaningfully reduce the demonstrated attack surface, making attacks less reliable and limiting their reach, even without completely eliminating the underlying problem,” he said.
Måløy also addressed the resolution that many analysts and consultants suggested, which is to change the nature of LLMs to fully isolate instructions from the data they operate on.
“Separating instructions from data may be part of the solution, but I think the distinction between data and instructions is not always clear in real-world workflows. For example, a user may ask an agent to arrange a business trip, requiring the agent to retrieve an email specifying the approved itinerary and a document containing the booking procedure,” Måløy said.
“My view is that the broader challenge is therefore not simply to prevent systems from interpreting external content as instructions, but to evaluate whether those instructions align with the user’s goals and the context in which the system is operating,” he added.
Problem hard to overstate
Mike Wilkes, enterprise CISO at Aikido Security, said it would be difficult to overstate the potential problems from this situation.
“This is a significant issue because it moves prompt injection from a single compromised interaction into a potentially self-propagating document integrity attack,” he said, noting that it is not a conventional worm that spreads automatically. A user or Copilot workflow must still bring an infected document into the model’s context. “But once that happens, the malicious instructions can reportedly alter business information, conceal themselves inside the resulting Word document and turn a legitimate internal file into the next carrier,” he said.
That, he pointed out, creates “a dangerous enterprise supply chain in which financial reports, contracts, policies and partner documents may inherit malicious behavior while retaining the trust associated with their legitimate authors and Microsoft 365 accounts.”
Not a new issue
The key issue behind this flaw is the fact that genAI tends to struggle with telling the difference between data that a user offers as input and the instructions the AI is to execute.
Flavio Villanustre, CISO for the LexisNexis Risk Solutions Group, observed that the same problem with mixed data and instructions occurred in databases decades ago, and turned into what we know as SQL injection attacks today. But, he noted, a few years later, parametrized binding for the database access layers was developed, which separates instructions, which are handled internally and safely, from data, which can come from untrusted sources. “The same needs to happen with LLMs and other AI,” he said.
Mike Leone, a VP/principal analyst at Moor Insights & Strategy, agreed.
“It’s hard not to chuckle a bit with this one. People have been asking whether data can give orders since SQL injections. We fixed that one by giving the database a way to tell an instruction from a value,” Leone said. “Thirty years later, we’ve built an entire category of software that can’t tell the difference at all.”
Tom Findling, CEO at Conifers.ai, also said he sees the data-vs-instructions issue as the biggest problem.
“Separating data from instructions is still unsolved at the model layer. The labs are making real progress, but not enough to be your only control,” he said, and stressed, “Microsoft doesn’t need to wait for that. It can slow propagation inside Copilot: stop hidden content from being written into new documents, surface hidden text before it’s used, preserve the trust level of what Copilot read and show all AI-generated changes.”
Fix needs industry agreement
However, Frank Dickson, group VP for security at IDC, argued that a fix needs industry agreement.
“A fundamental fix requires the industry to converge on the same architectural change, which isn’t happening anytime soon given the incentives and the fact that the major players are far from besties,” he said, pointing out that Instruction-data separation would need to be baked in at the model or platform level across every major vendor. “None of them are rewarding that work commercially right now, so treat that as a multi-year research problem, not something a CISO should wait on.”
Leone disagreed, arguing that a single vendor can do quite a bit, at least for its own customers.
“Microsoft has said publicly since last year that indirect prompt injection can’t be fully prevented, and I give them credit for saying it out loud rather than pretending otherwise,” Leone said, noting that the industry doesn’t have to move in lockstep. “This attack lives inside one product,” he said. “If Microsoft hardens the Copilot document path, Copilot customers get safer whether or not anyone else moves.”
Dickson added that another critical problem with this vulnerability is that the carrier documents are created by legitimate employees using legitimate tools, “so there’s no obvious point of origin once the payload starts moving; an organization could be laundering corrupted financial figures through its own workflow for weeks before anyone notices.”
Tactics that might help
That said, experts had some specific suggestions for CISOs trying to negate this problem.
IDC’s Dickson said, “the most useful lever right now sits outside the model entirely, in how much untrusted content Copilot is allowed to pull into a session without a human choosing it.” He noted that enterprises can turn Copilot’s auto-discovery behavior off or restrict it, and require humans to explicitly select the documents they feed it. “That alone closes off one of the two ways the attack gets a foothold,” he said.
IDC also suggested a visible diff or redline of anything Copilot changes in a financial or otherwise consequential document, and the requirement for a human to approve the changes. “That’s not a technical fix, it’s a workflow one, and it’s available today,” Dickson said.
He said that IT should also track where content came from and what was touched by an AI system, in metadata that travels with the document. “It doesn’t stop the injection, but it means when something does get through, an organization can actually trace how far it spread instead of discovering a corrupted number three reports later with no way to reconstruct the chain,” he said.
However, one expert, Tyler Reguly, Fortra’s associate director of security R&D, said that he didn’t see this vulnerability having a meaningful impact because it feels like what he called a “laboratory vulnerability.”
Normal enterprise workflows don’t encompass a number of the necessary steps for compromise, he said, pointing out that, if anything, people are trained not to download Word documents. Plus, he noted, looking at the blurred example in the report, the malicious document contained an additional apparently blank page which held the concealed prompts in white text.
“That page is going to send up warning signs for me if someone even managed to convince me to download an external Word document,” he said. So, overall, “this feels like it requires a perfect storm.”
This article originally appeared on CSOonline.
Read the full article here
