Skip to Content

Prediction 01

Assistant-Driven Intrusions

Enterprise AI Agents Will Accelerate Post-Compromise Intrusions

In 2027, attackers who compromise an account are highly likely to co-opt approved enterprise AI assistants as a stealthy alternative to bulk data theft, querying these tools to locate high-value records while staying beneath mass-download detection thresholds.

~12 Min Read

Over the past two years, AI has reshaped enterprise workflows. Companies are deploying agentic AI assistants across departments to boost efficiency and productivity, but doing so requires giving these agents permission to access sensitive systems and data. Those exact permissions are what threat actors go after.

Adversaries traditionally default to broad data theft because they lack the time to manually search unfamiliar environments, even though bulk exfiltration risks triggering security alarms. When an approved AI assistant is available, an attacker can use it to locate the most sensitive assets directly. By keeping data movement focused, threat actors can evade mass-download detection rules and blend their retrieval into ordinary workplace traffic.

The Rise of the Enterprise AI Assistant

Organizations have embedded copilots, assistants, and agents into nearly every part of their work, from email, calendars, and documents to code and workflow automations. In most enterprises, that’s now the normal way of working. But each of those deployments is connected to sensitive data and business processes by design, because an assistant without that access would work from poor data and produce low-quality results.

What AI Assistants Reach in the Enterprise
Email, calendars, documents, code and workflow automations, each connected to sensitive data and business processes.
Email
Calendars
Documents
Code
Workflow automations
Embedded copilots, assistants, and agents
Sensitive data and business processes
Illustrative. Surfaces are those named in the paragraph above.

Most organizations likely started with generative AI (GenAI) chatbots that answered questions and drafted emails. As AI capabilities grew exponentially, enterprises began deploying agentic AI assistants. That’s a major step up for employees. Instead of just responding to a prompt, an agentic assistant can be given an objective and expected to complete it. It plans the steps, calls the tools it needs, checks whether each step worked, and adjusts without being prompted again. In practice, it can create a sharing link, send a message, move a file, or change a permission. For an adversary with access to a compromised account, these built-in capabilities hand them an authorized, ready-made toolkit for post-compromise reconnaissance and data staging.

From Answering to Acting
An agentic assistant is given an objective and completes it, planning, calling tools, checking and adjusting without further prompts.
GenAI chatbot

Responds to a prompt

One question in, one answer out
Agentic assistant

Completes an objective

  1. 01
    Plans the steps
  2. 02
    Calls the tools it needs

    Create a sharing link, send a message, move a file, or change a permission.

  3. 03
    Checks whether each step worked
  4. 04
    Adjusts without being prompted again
Illustrative. Actions are those named in the paragraph above.

Racing the Clock: The Attacker’s Dilemma

Post compromise, threat actors operate against the clock, aware that their access window may close the moment an account is flagged or an active session is terminated. While some extortion operators and specialized access contractors perform preliminary enumeration to identify and target specific high-value repositories such as financial or legal SharePoint sites, manual sifting remains too slow under time pressure. As a result, attackers still default to volume within those targeted repositories, rapidly exfiltrating large batches of data to parse offline before security teams intervene.

In one incident we investigated, that approach produced automated bulk retrieval of over a million files and close to a million customer account records exported from a separate platform through a single query. The alert that fired was a mass-download rule, so volume is what gave the attacker away.

One incident we investigated
1,000,000+
files retrieved through automated bulk access
~1,000,000
customer account records exported from a separate platform through a single query

Co-opting an approved enterprise AI assistant opens an alternative path with distinct trade-offs. Rather than replacing traditional exfiltration playbooks, it gives attackers added flexibility.

First, it substantially reduces data volume and lowers detection risk. Sifting specific, high-impact records before exfiltrating them moves far fewer bytes across the network, which cuts the odds of tripping threshold-based alerts.

Second, the activity blends into authorized business operations. Because the queries originate from a sanctioned enterprise application using permissions it already holds, the resulting audit logs closely resemble regular employee activity.

Third, it accelerates discovery in an unfamiliar network. An adversary entering a new environment traditionally spends valuable time running noisy discovery commands to map directory structures. An AI assistant has already indexed those repositories, allowing the intruder to locate critical systems and documents immediately.

Many adversaries will still likely choose bulk theft, particularly when operational speed outweighs stealth or when an extortion model demands sheer volume. The presence of an enterprise AI assistant doesn’t eliminate traditional smash-and-grab attacks. It equips threat actors with a quieter, low-volume alternative designed to bypass the volume thresholds defenders rely on most.

From Smash-and-Grab to Ask-and-Extract

Attackers use the smash-and-grab playbook after a compromise because they’ve landed in an unfamiliar environment with limited time and no maps. But an attacker who has compromised an account with an agentic AI assistant can sidestep that problem relatively easily. With an agentic AI assistant in tow, the attacker can use this capability as an on-demand research assistant, helping it analyze compromised environments, explore potential post-compromise techniques, and troubleshoot or adapt tools to specific operational constraints.

A case ReliaQuest investigated shows how much power these assistants hold through their permissions. Several employees at one organization used an approved AI assistant to reach the corporate document store through a pre-authorized application. They retrieved files programmatically rather than through a browser and extracted up to a thousand documents per session. It later emerged that the users were several senior executives pulling financial models, regulatory filings, and accounting ledgers.

Although this was legitimate business activity rather than a breach, it illustrates the power of standing permissions. An attacker using this exact access path could retrieve high-value corporate documents at scale, and the resulting telemetry would be indistinguishable from the routine operations of an approved enterprise tool.

What's the Impact on Enterprises?

As enterprises wire AI assistants into the systems that run the business, the access that makes the assistants useful is the same access attackers prize after a compromise.

Enterprise AI assistants turned against their owners pose three risks:

  1. Reconnaissance Time Shrinks to Minutes

    Post compromise, attackers often need answers to critical questions, such as where the domain controllers are, which accounts carry admin privileges, and what security tools are running. This costs the attacker precious time and raises the risk of getting caught before they reach their objectives. Using an AI assistant post compromise would help attackers quickly analyze system and network data, summarize directory structures and configuration files, and make sense of the target’s tech stack. As a result, attackers will move faster and with greater precision. Organizations should review what system and network data their AI assistants can access and query and treat that visibility as a potential attack surface.

  2. Attackers Take What Hurts the Most

    Attackers often default to exfiltrating large datasets because they may lack the time to manually search complex file repositories. An enterprise-approved AI assistant changes that dynamic by helping the adversary locate and synthesize high-leverage data directly, so they can steal critical assets while keeping network volume low enough to avoid tripping mass-download alarms. By filtering for the most sensitive files, attackers can run focused data theft that blends into ordinary application traffic. Defenders should look beyond simple download volumes and monitor behavioral anomalies—such as an approved assistant suddenly querying repositories it has never touched or an account issuing requests against infrastructure far outside its established operational profile.

  3. Credential Harvesting and Lateral Movement

    An enterprise agentic AI assistant is likely connected to the tools employees use every day, including code repositories, secret managers, internal knowledgebase documents, and other cloud infrastructure. Attackers with access to a compromised account can use the assistant to query those systems, surface API keys stored in a repository, pull credentials saved in a shared document, identify service accounts with elevated privileges, and map out which systems those accounts can reach. For defenders, these actions will generate logs that look identical to a user doing their job, so suspicious behavior may be missed or closed as routine. To counter this risk, organizations should audit and restrict the systems an assistant can query, enforce least-privilege scoping so assistants can’t access central secrets stores, and monitor for sudden spikes in cross-platform credential queries.

Defend Against Intrusions Driven by AI Assistants

Countering the Threat with ReliaQuest

Much like traditional service accounts, an enterprise AI assistant holds broad standing privileges that security teams can’t easily revoke without breaking the business workflows it supports. Defenders can’t simply strip away that standing reach, but they can govern the order in which those permissions are exercised and intervene before bulk data extraction finishes.

Post compromise, an attacker who piggybacks on an approved assistant inherits legitimate permissions the organization granted deliberately. In isolation, no individual step appears suspicious, as the assistant receives consent, queries enterprise data stores, and retrieves documents. The assistant can even serve as an inadvertent guide. When a direct query encounters an access restriction, its conversational design frequently prompts it to propose alternative pathways, such as summarizing files it can’t directly export or pointing the user toward related accessible tables. The intrusion only becomes visible through its operational sequence and velocity, typically moving from broad reconnaissance to volume collection. Defeating this threat requires real-time correlation across seemingly benign events, streaming detection, and automated containment.

Real-Time Correlation with GreyMatter Transit

GreyMatter Transit is ReliaQuest’s security data pipeline with detection at source and in transit. Rather than waiting for log batches to land in a SIEM, use Transit to detect on data in motion, cutting out ingestion delays. Data gets enriched with threat intelligence as events stream between tools, so threats surface earlier, with no waiting for data to land or for manual enrichment.

For this threat scenario, Transit’s correlation engine would perform multi-event detection in transit:

  • Sequence: Transit tracks related actions across time to spot when a consent grant is followed by suspicious searches, even amid normal user activity.
  • Volume: Transit tracks event frequency per entity in real time, firing the moment retrieval counts hit configured thresholds. An approved AI assistant downloading a couple of files is likely normal activity. But an agent suddenly pulling hundreds of documents in a few minutes’ triggers immediate containment.

As Transit correlates sequence and volume while data is still streaming, detection completes while sessions are live, giving security teams the speed needed to disrupt threat actors before exfiltration finishes.

Detection Rules

The detection rules below cover the three points where the chain becomes visible.

Full Access Granted to an OAuth Application

Triggers on single Entra ID events when applications are consented to with permission to act across tenants as themselves rather than on behalf of signed-in users. In an assistant-driven intrusion, adversaries may attempt to expand the reach of an AI integration by granting it broad application-level access or connecting malicious third-party plugins to the tenant. This is the earliest control point available, catching attackers who phish consent for applications of their own before any data moves.

SharePoint Site Enumeration

Flags site-wide and wildcard search queries against SharePoint Online from unmanaged devices using scripted clients. It catches the reconnaissance that attackers perform when they don’t know what the files hold.

Mass File Downloads from File Share

Triggers on high-volume retrieval of unique files by a single account inside a short window. This rule catches the shift from reconnaissance to bulk exfiltration, the moment an agent begins pulling documents at machine speed. Catching anomalous volume spikes before files leave the network lets defenders cut off the session before an automated harvest finishes.

  1. Ingest Cloud Audit Logs Before Anything Else

    Enable unified audit logging across document repositories, customer relationship management (CRM) platforms, and collaboration tools before relying on endpoint detection. Endpoint telemetry can’t capture an intrusion that occurs entirely in cloud-to-cloud workflows, making software-as-a-service (SaaS) and repository log ingestion the prerequisite for visibility.

  2. Baseline Every Approved Assistant, per Application

    Record which applications retrieve at volume, which accounts invoke them, in what band, during which hours, and against which sites. Restrict assistant access to secrets managers, code repositories, and administrative tools unless an explicit business use case requires it. Inventory which assistants can act rather than only answer, and scope their permissions below the invoking user's where the platform allows. Without that baseline, an analyst will likely have little or no basis for treating the next mass-download alert differently from the last.

  3. Correlate Identity and Behavioral Signals Independent of Volume Thresholds

    Don’t rely on mass-download alarms as the primary trigger, because attackers actively keep extraction volumes low to evade threshold-based detection. Instead, correlate subtle behavioral cues across the entire session, flagging when an assistant queries repositories it has never touched before, or when access originates from an unmanaged device. Also, verify application consent events and recent authentication-method changes within the same session before dismissing lower-priority alerts, and enforce strict device compliance policies for any connection attempting to reach repositories holding regulated data.

Live Event

See GreyMatter Live

Agentic Defense in Action: Contain Threats in Minutes with GreyMatter