// 1 ZERO-DAY · 2 CVE · 2 EXPLOIT IN THE LAST 24H→
The Fulcrumsec extortion group released "Stage 2" of its attack on Novo Nordisk on August 13, 2026, publishing what it claims are 30 Hugging Face AI models, 70 datasets, and 0.5 terabytes of proprietary cell painting microscopy images. The tranche totals 1.05 TB, bringing the alleged total haul to 1.3 TB of data stolen from the Danish pharmaceutical giant specializing in diabetes and obesity therapies.

The Fulcrumsec extortion group released "Stage 2" of its attack on Novo Nordisk on August 13, 2026, publishing what it claims are 30 Hugging Face AI models, 70 datasets, and 0.5 terabytes of proprietary cell painting microscopy images. The tranche, which the group quantifies at 1.05 terabytes, brings the alleged total leak to 1.3 terabytes of data stolen from the Danish pharmaceutical company specializing in diabetes and obesity therapies.

The publication marks a qualitative shift in the threat landscape for the Life Sciences sector: for the first time, an extortion actor claims not only the theft of clinical and proprietary data, but the complete exfiltration of an industrial AI ecosystem, including the training artifacts that determine its predictive capability.

Key Takeaways
  • Fulcrumsec published "Stage 2" of the Novo Nordisk leak on August 13, 2026, comprising 30 Hugging Face models, 70 datasets, and 0.5 TB of microscopy images, totaling 1.05 TB according to the group
  • Initial access occurred via secrets exposed in client-side JavaScript on two separate Novo Nordisk subdomains, according to the threat actor's statement
  • Novo Nordisk confirmed unauthorized access to "a limited number of internal IT systems" on June 11, 2026, and later acknowledged the external copying of "some non-public data, including personal data"
  • The group claims to hold PII for 46,843 Danish healthcare professionals, 506,007 clinical investigators, and 2,852 clinical trial staff, stating it has "withheld per agreement"

The Technical Vector: Frontend Secrets as Entry Point

Fulcrumsec describes a recurring design failure. According to the group's post, cited in Marianne Kolbasuk McGee's article on HealthInfoSec/BankInfoSecurity: "We gained initial access through secrets left in client-side JavaScript on two separate unrelated Novo Nordisk subdomains - two completely different teams, two different applications, the same elementary mistake made twice."

The group's statement highlights a specific vulnerability in the frontend development pipeline: the compilation of JavaScript bundles that include credentials or tokens with access to cloud services, making them retrievable by any page visitor. Fulcrumsec underscores the lack of monitoring with a reproachful tone: "a $400 billion corporation with a dedicated cybersecurity division cannot be bothered to monitor their frontend bundles."

The group adds a significant chronological element regarding undetected persistence: "That they could not detect unknown IPs raiding their cloud services for weeks and months before responding - or never detecting us at all, in the case of their HuggingFace and Okta accounts." This assertion, not independently verified, if confirmed would indicate a dwell time measurable in months for certain accounts.

What Stage 2 Contains: From Source Code to Model Weights

The first leak tranche, published in June, had according to Fulcrumsec exposed "compounds, their manufacturing recipes, and their source code." Stage 2 expands the scope to the enterprise machine learning infrastructure. The group states the new tranche includes "30 Hugging Face AI models, 70 datasets and half of a terabyte of proprietary cell painting microscopy images."

Cell painting is a high-content microscopy technique that uses six fluorescent channels to capture cellular architecture, generating profiles usable for predicting pharmaceutical compound activity. The theft of these datasets, combined with the models trained to interpret them, represents a two-level compromise: the raw data and the inferential capacity developed to extract value from it.

Fulcrumsec distributed the materials "via torrent for now, with additional mirrors coming soon," according to a direct quote from its post. The group also claims a total of 1.3 terabytes of stolen data, contrasting itself with another actor, TheUSERS007, who published a separate claim in June: "Even if real, they are claiming something like 18 GB; we have 1.3 TERABYTES."

"Our Stage 1 release gave the world Novo's compounds, their manufacturing recipes, and their source code. Stage 2 gives the world the machines and datasets used to create new ones." — Fulcrumsec, darkweb post cited by HealthInfoSec

The Rivalry with TheUSERS007 and Withheld PII

The incident presents an uncommon complication: the overlap of two independent claims. In June 2026, TheUSERS007 communicated its own Novo Nordisk breach to Databreaches.net, creating uncertainty over the provenance of the initial access. Fulcrumsec contested the priority and extent of the rival compromise, explicitly linking its own action to the company's official disclosure: "they are not claiming the patient or doctor PII, which precipitated Novo's disclosure; that was us."

Alongside the publication of research materials, Fulcrumsec claims to hold personal information on three categories of individuals: 46,843 Danish healthcare professionals, 506,007 clinical investigators, and 2,852 named clinical trial staff. The group states it has "withheld per agreement," leaving open the question of the nature of this alleged agreement and its counterparties.

Novo Nordisk did not respond to ISMG requests for comment on Fulcrumsec's specific claims, limiting itself to reiterating statements already made in June. The company has neither confirmed nor denied the data volumes, the described access vectors, or the presence of AI models within the scope of the copied materials.

Why It Matters

The dossier presents significant limitations that affect the assessment of the incident. There is no independent or forensic confirmation of the authenticity of the data published in Stage 2, nor has Novo Nordisk verified the claimed volumes of 1.3 terabytes, 30 models, and 70 datasets. The primary source reports threat actor statements without the possibility of cross-verification.

The discrepancy regarding the initial access vector remains unresolved: the primary source identifies "secrets left in client-side JavaScript," while the tertiary source Tech-Insider introduces a "GitHub token leaked in March" not present in the primary sources.

It is not verified that the materials are actually publicly accessible via torrent, nor that the buyer interest in China claimed by Fulcrumsec is real. The group states it rejected offers to "open source" the data, but this narrative is not confirmable. The status of regulatory investigations (GDPR, Danish authorities) is not detailed in the dossier.

The statement by Lee Kim, founder of Keytera and former HIMSS executive, reported by the primary source, places the incident in a broader legal framework: companies whose IP is stolen in a cyberattack can seek injunctions, takedowns, recipient notifications, civil and criminal actions, and measures to block the import of products developed with stolen IP. Her quote, however, refers to a general horizon of legal remedies, not to an action undertaken by Novo Nordisk.

Frequently Asked Questions

Has Novo Nordisk confirmed the theft of AI models?

No. The company confirmed in June that "some non-public data, including personal data" was copied externally, but has not specified the nature of the materials nor responded to ISMG requests regarding Stage 2.

Is the leak verified as authentic?

No. The authenticity of the data published in Stage 2 has not been confirmed by independent sources or by Novo Nordisk itself. The volumes and composition of the leak are Fulcrumsec's claims.

What is the strategic value of the stolen AI models?

According to the dossier's editorial angle, the trained models represent "industrial predictive capability" that goes beyond raw data, with implications for intellectual property and competition in the pharmaceutical sector. This interpretive reading is not, however, supported by official statements from Novo Nordisk.


Sources

Sources and references
  1. bankinfosecurity.com
  2. tech-insider.org
  3. cyberera.com.ng