Storage-Level Exfiltration Detection Without Alert Fatigue

Storage-Level Exfiltration Detection Without Alert Fatigue

Table of Contents

Restoring from a clean backup ends the encryption problem. It does not end the extortion problem, because the data has already left the building.

Storage-level exfiltration detection finds data theft from the access patterns that volumes, file systems and object stores record. Examples are abnormal read volume, unusual access breadth and off-hours bulk reads. It works before or alongside encryption. The need is measurable. Dual extortion ransomware, where threat actors encrypt systems and steal data at the same time, accounted for 70% of all ransomware claims in 2025, according to the Coalition 2026 cyber claims report. In the fastest cases it investigated, the Unit 42 2026 global incident response report found that attackers needed just 72 minutes to move from initial access to data exfiltration.

This guide explains what the storage layer can reveal and where that signal misleads. It also shows how to run it without burying the security operations center (SOC) in alerts. Limits are the core of the evaluation. A team that knows what a read pattern cannot show makes a better purchase and a better deployment. Ransomware detection that watches only for encryption covers half of the attack.

Why Ransomware Detection Must Start Before Encryption Begins

CISA ransomware guidance defines double extortion as encrypting data and stealing it. The same guide adds that “in some cases, malicious actors may exfiltrate data and threaten to release it as their sole form of extortion without employing ransomware.” Theft is staged ahead of the visible event. That staging creates a window in which a defender can act.

Current incident data points the same way. In Cybersecurity Dive coverage of the Arctic Wolf 2026 Threat Report, data-only extortion rose from 2% to 22% of the cases Arctic Wolf handled between November 2024 and November 2025. That is an elevenfold increase. Help Net Security coverage of the Picus Red Report 2026 reports a 38% year-over-year decline in Data Encrypted for Impact, MITRE technique T1486, from 21.00% in 2024 to 12.94% in 2025. Attackers are shifting weight from locking data to quietly taking it.

Bar chart of data-only extortion share of Arctic Wolf cases: 2% in November 2024 against 22% in November 2025.
Data-only extortion rose from 2% to 22% of Arctic Wolf cases between November 2024 and November 2025, per its 2026 Threat Report.

Encryption has not disappeared. The Sophos State of Ransomware 2026 report found that 56% of attacks still succeeded in encrypting data, up from 50% the year before. Defenders need to see both halves. Missing the theft half is costly: the Coalition report found that ransomware attacks involving data theft were more than two times more expensive than other ransomware variants.

Immutable backups remove the encryption lever, not the leak lever. A write-once copy cannot be encrypted or deleted by an intruder, as the StoneFly guidance on immutable file-level WORM explains. But it does nothing about a copy of the data already sitting on an attacker’s server.

What Storage-Level Exfiltration Detection Can See

Storage systems record who touched what, how much and when. That includes read throughput per volume, share or bucket. It includes file and object open counts per identity, the breadth of directories reached, authentication events, session timing and file change records. None of this requires inspecting packet contents. That is why it holds up even when traffic is encrypted.

This view matters most for stolen-credential theft. The Sophos report found that 79% of ransomware attacks started with an identity-based approach. The Unit 42 report states that identity weaknesses played a material role in nearly 90% of its investigations. A stolen credential produces a normal-looking sign-in and permitted traffic. It does not produce a normal read pattern. Take an account that usually opens a few hundred files from one department share. It suddenly reads tens of thousands across every share. The storage layer records that change, and the network layer cannot judge it.

MITRE frames the goal plainly: the MITRE ATT&CK exfiltration tactic covers techniques adversaries use to steal data from your network. For object stores, the MITRE guidance on data from cloud storage lists detection ideas built on access-pattern anomalies. One example is a spike in object access from a new identity, followed by exfiltration to external addresses. Storage sees the first half of that sequence directly.

Detection of this kind rests on behavioral baselining. That means comparing current activity with the normal pattern for each user, volume and process. It also rests on retained logs. NIST draft guidance on log management describes log management as the process for generating, transmitting, storing, accessing and disposing of log data. Without that discipline, a baseline has nothing to learn from.

Storage-Level Exfiltration Detection and the False-Positive Problem

A theft pattern at the storage layer usually combines four traits. There is a sudden jump in read volume. Access spreads across many directories. The identity does not normally touch that data. The timing is off-hours. The CISA advisory on Gunra ransomware shows the pattern in the field. The actors worked late at night and early in the morning, between 10:00 p.m. and 06:00 a.m. They used 7-Zip, RClone and FileZilla for collection and exfiltration. In one case they moved tens of terabytes to a file-sharing service. CISA recommends auditing file creation records, particularly archives, to find this activity.

Four cards: sudden jump in read volume, access spread across directories, identity outside its normal data, and off-hours timing.
Four read-side traits that, taken together, point to data theft at the storage layer.

The difficulty is that normal work produces the same shapes. Common look-alikes include:

  • Backup jobs, which read entire volumes on a schedule.
  • Replication, which copies large data sets between sites.
  • Search indexing, which touches every file in a share.
  • Antivirus and integrity scans, which open files across the whole namespace.
  • Migrations and analytics batch jobs, which move or scan large data sets in a few hours.

Each of these can trigger a rule that was written for theft. When that happens, analysts learn to dismiss the alert. The dismissed alert is the one that matters on the day of a real incident. The cost is real. The Picus Blue Report 2025, as reported in Help Net Security coverage of Picus research, found that 54% of attacker activity is logged but only 14% generates an alert. The gap sits between the logged activity and the alert, so tuning matters more than collection.

Bar chart: 54% of attacker activity is logged, while only 14% generates an alert.
The Picus Blue Report 2025 found that 54% of attacker activity is logged but only 14% generates an alert.

Four tuning practices address most of the noise. First, baseline for a full business cycle so that month-end jobs count as normal. Second, allow-list known service accounts, job schedules and source and destination pairs. Third, tier severity so that a single anomalous read raises a low-priority event rather than a page. Fourth, require correlation before paging. For example, page only on a bulk read plus an unusual identity plus an off-hours window.

What Storage-Level Exfiltration Detection Cannot Tell You

Storage-level detection cannot see the destination, the protocol, or whether data left the network at all. A large read followed by a copy to another internal share looks the same as a read followed by an upload to an external host. Only the network layer can tell them apart.

It also cannot see intent or content. A read pattern shows that an identity touched many files. It does not show whether those files held source code or lunch menus. It does not show whether the person was an analyst finishing a project or an intruder. Context has to come from identity, endpoint and data loss prevention signals.

Slow or small-subset theft can stay inside a baseline. MITRE documents that an adversary may exfiltrate data in fixed-size chunks instead of whole files or limit packet sizes below certain thresholds, as described in the MITRE entry on data transfer size limits. The same logic applies to reads. An attacker who copies one directory a night for a month produces no spike. Detection also cannot cover data that never touched the monitored storage, such as files held only on laptops or in a software-as-a-service application.

Immutability has a similar boundary. Features such as object lock for protected data keep stored copies from being altered. They do not stop a valid identity from reading them. Detection and immutability answer different questions, and a sound design uses both.

Finally, a detection is only the start of the response. NIST SP 800-61 Revision 3 places incident response inside broader cybersecurity risk management. A storage alert is useful when it feeds that process through a security information and event management (SIEM) platform, where it can be joined with other signals.

Can Storage-Level Detection Replace Network or Endpoint Tools?

No. Storage-level detection adds a vantage point that network and endpoint tools lack, namely what was read and by whom. But it cannot see destination, process lineage or data content. Treat it as one layer in ransomware detection. Join it to network detection and response and endpoint detection and response through SIEM correlation.

How StoneFly IMS365 Supports Ransomware Detection and Response

StoneFly IMS365 is the Infrastructure Management and Security Platform with Agentic AI. It gives enterprise teams a single view of management and security across multi-site and hybrid environments, including Proxmox, VMware, and Hyper-V hypervisors, VMs, Windows and Linux servers, optional Veeam and Rubrik environments, and storage. That single view is what the operating model in this article needs, because a storage alert is only useful when it can be joined with server, hypervisor and backup context.

IMS365 uses agents and syslog monitoring to identify threats, and it can send threat alerts to an administrator Slack channel when a configured threshold is reached. Reports can be generated on demand or scheduled for delivery by email, Slack or Telegram. Its Essentials tier covers this monitoring, threat visibility and reporting, and administrators investigate and remediate manually. Which signals are collected depends on the agents and connections you configure.

The Enterprise tier adds InfraAI365, an AI command center. Administrators can ask questions about a threat detection, receive remediation guidance, and run system playbooks that can coordinate storage and hypervisor actions. Examples include a volume snapshot, an approved access restriction, or isolating a Proxmox/VMware/Hyper-V VM, followed by a check of the reported result. Playbooks can also orchestrate connected SOAR and SIEM tools. This is where the tuning advice above becomes operational:

  • An action policy decides which steps run directly, which need human approval and which an administrator performs manually.
  • Dangerous defensive actions require explicit approval, and the record captures the approver’s identity, source IP address and time.
  • A response preview mode logs what a playbook would do before live execution is enabled.
  • After an action, IMS365 checks connected systems for the reported result and remaining threat signals, and an administrator authorizes release from isolation.

When a playbook calls for it, InfraAI365 can start a forensic backup automatically, and it can coordinate VM and storage snapshots before a high-impact action to preserve evidence. The backup engine BKD365 connects Veeam and Rubrik environments to the same view, which Enterprise playbooks need for backup operations. These actions depend on the configured connection, permissions and target system, and they do not guarantee rollback or complete threat removal.

Running Storage-Level Exfiltration Detection Without Alert Fatigue

How you run it matters more than any single rule. A workable sequence for storage-level exfiltration detection looks like this:

  1. Baseline for a full business cycle before enforcement, so periodic jobs are learned as normal.
  2. Allow-list backup, replication, indexing and scan identities together with their schedules.
  3. Tier severity, and page only on correlated signals such as bulk reads plus an atypical identity plus an off-hours window.
  4. Route storage events to the SIEM and join them with network and endpoint events.
  5. Pre-approve containment actions that are safe to automate, such as isolating a session, and require human approval for the rest.
  6. Review false positives monthly and feed each one back into the allow-list and the baseline.
Six cards: baseline fully, allow-list known jobs, tier and correlate, route to the SIEM, pre-approve containment, review monthly.
The six-step operating model for running storage-level exfiltration detection without alert fatigue, from baselining to monthly review.

Before choosing a platform, put the same questions to every vendor:

  • Which telemetry does the platform analyze: storage I/O, file activity, authentication events, processes or network events?
  • How are baselines built, and how long is the learning window?
  • Which response actions exist, and which can be limited to policy-approved cases?
  • What evidence is retained for forensic work, and for how long?
  • How does it integrate with the SIEM we already run?

Ask for evidence from your own environment rather than a quoted detection rate. Rates depend on the baseline, the workload and the configuration.

Conclusion: Detect Early, Respond by Policy, and Keep Ransomware Detection Tuned

Backups answer the encryption half of double extortion. The theft half needs a signal that appears before the ransom note. The storage layer provides one that is tied to identity and to real read behavior. Storage-level exfiltration detection will not show destination, intent or content, and it can miss patient, low-volume theft. It belongs beside network and endpoint detection, not in place of them.

The decision for a CTO or storage architect is how to combine the layers and how to protect analyst attention. Baseline fully, allow-list known jobs, tier severity, correlate before paging and automate only the containment steps you have approved in advance. Done well, storage-level exfiltration detection turns the access logs you already generate into an early warning.

To see how IMS365 and integrated threat detection and response fit your storage platform, talk to StoneFly about your storage platform.

Related Products

StoneFly DR365V Veeam Ready Backup & DR Appliance

Unified Storage and Server (USS™) Hyperconverged Infrastructure (HCI)

Unified Scale-Out (USO™) SAN, NAS, and S3 Object Storage Appliance

Subscribe To Our Newsletter

Join our mailing list to receive the latest news, updates, and promotions from StoneFly.

Please Confirm your subscription from the email