Artificial Intelligence & Machine Learning,
Cybercrime,
Fraud Management & Cybercrime

In addition, Russian Hackers Exploit Outlook Flaw, Coca-Cola Restarts Fairlife Production

Anviksha More
• July 30, 2026

Every week, Information Security Media Group (ISMG) compiles noteworthy cybersecurity incidents from around the globe. This week, Sam Altman, the CEO of OpenAI, commented on the possibility that the company’s AI models might have inadvertently accessed other organizations’ systems. His remarks were made shortly after OpenAI revealed a breach involving Hugging Face’s code repository.

Sam Altman suggested during an address at Capitol Hill that it was plausible for OpenAI’s models to have hacked other systems, stating, “I mean there could be, yeah.” This remark followed an update from OpenAI about their breach of Hugging Face, indicating they found instances where their models had used publicly exposed credentials from various other services.

OpenAI detailed that four accounts on unnamed services had been accessed during the Hugging Face incident. Among these, one account served as an outbound relay and staging point, while another was for data storage. The other two accounts were accessed only in a limited manner and did not contribute to the attack against Hugging Face. OpenAI assured stakeholders that they would notify service owners directly and noted they had not observed any indication of broader implications for these services.

According to a report by Reuters, one of the compromised entities is believed to be Modal Labs. The firm’s CTO, Akshat Bubna, confirmed that the models exploited the code written by a customer hosted on Modal’s cloud infrastructure.

A Russian nation-state cyber-espionage group has resumed operations after a period of dormancy, launching a campaign utilizing a newly patched vulnerability in Microsoft Outlook Web Access (OWA) to infiltrate organizations and sustain long-term access to their email systems. This activity, attributed to the group known as TA488, commenced on July 22 and targeted government agencies in the U.S. and Europe, as well as sectors such as telecommunications, financial services, hospitality, and aerospace.

Attackers took advantage of a cross-site scripting vulnerability identified as CVE-2026-42897, which affects the on-premises Microsoft Exchange Server. The simplicity of the exploit, where merely opening a malicious email in a vulnerable OWA client triggers the attack, is concerning since it requires no additional user engagement.

Proofpoint’s findings indicated that the attackers used seemingly innocuous emails, luring targets with subjects related to supply chains, gas markets, and tourism, thereby evading suspicion. The implemented malware, described as OWAReaper, functions within the OWA reading pane without leaving physical traces, enabling it to harvest user credentials and alter original emails to erase traces of the exploit.

The malware established long-standing access by abusing Exchange mailbox permissions, ensuring that even after changes in passwords or reconditioning of infected devices, the attackers could retain access. The malware communicated with command servers using GitHub commit messages and email, delivering stolen data via HTTPS with DNS tunneling as a secondary method.

Coca-Cola’s Fairlife milk division has mostly resumed its production following a ransomware attack that had halted operations at four of its U.S. plants earlier this month. The beverage corporation announced significant advancements in restoring impacted systems, allowing production to restart at the affected facilities.

The company first disclosed the security incident on July 16, noting that an unauthorized entity had gained access to parts of Fairlife’s technological environment. Although the attack temporarily suspended production, Coca-Cola asserted that product quality and safety were never compromised.

The company further stated that the disruption is unlikely to notably affect sales, assuring investors that existing inventory has maintained Fairlife product availability at retail locations. It’s noteworthy that Coca-Cola had acquired Fairlife from Select Milk Producers in a substantial deal worth around $7 billion back in 2020.

The U.K. Department for Education reported a cyberattack that has led to the exposure of approximately 607,000 educational records. This breach pertains to its customer helpdesk and the Turing Scheme, which is an international education funding initiative. The department emphasized that the compromised information was restricted to customer service contact details of various individuals and organizations, stating that other systems or data were not affected.

Media outlets reported that the breached data includes names and email addresses of head teachers, university personnel, and government officials. The cybercrime group known as ExfilSquad took responsibility for this incident, claiming to have stolen around 600,000 records from the customer helpdesk and an additional 7,000 from the Turing portal, which included sensitive information such as names, email addresses, telephone numbers, and job titles.

A cyberattack on Angola’s largest telecommunications provider disrupted mobile voice, data, and internet services nationwide, coinciding with the government’s sale of a 15% stake in the company on the national stock exchange, a significant portion of the quarter stake that was seized in 2020.

The attack was detected shortly after 2 a.m. local time, with technical and cybersecurity teams mobilized to restore services. However, details regarding the nature of the attack and a specific timeline for the complete resumption of operations remain undisclosed.

Network data analyzed indicated that the disruption likely stemmed from an internal system failure rather than an external attack or a distributed denial-of-service incident. The outage notably affected point-of-sale payment systems operating on Unitel’s network, causing significant disruptions for businesses and digital services across Angola.

A large-scale credential stuffing campaign targeted SonicWall VPN and firewall accounts, compromising 92 user accounts across 30 organizations in a span of less than two days. This attack began on Saturday and persisted for around 41 hours before abruptly ceasing on Monday.

Research from Huntress revealed that the campaign was broad in nature, appearing to target any internet-facing SonicWall device rather than concentrating on specific sectors or organizations. Notably, researchers observed no post-compromise interactions, suggesting that the attackers may have been pre-positioning access for future operations.

The precise source of these intrusions is still under investigation, but initial findings indicate that the attackers began with successful, authorized logins, likely utilizing previously stolen credentials. The incident closely follows a series of attacks against SonicWall products, amplifying concerns about the security of their systems.

In a significant development, Russia has issued an arrest warrant for Telegram’s founder, Pavel Durov, accusing him of facilitating terrorism through the messaging platform. This escalation highlights the ongoing tension between Russian authorities and the tech entrepreneur. The Federal Security Service claims that Telegram failed to remove content exploited by extremist groups to coordinate sabotage and cyber fraud within the country.

The severity of the allegations includes charges of aiding terrorism, which could result in a life sentence if Durov is convicted. In response to the accusations, Telegram posted a defiant image of Durov, signifying their resistance to the charges.

Security researchers have unveiled a proof-of-concept attack that can convert Microsoft Copilot for Word into a self-propagating AI “worm” by embedding hidden prompts within Word documents. This attack exploits prompt injection, utilizing covert instructions concealed as white text or within invisible document elements.

Although invisible to users, Microsoft Copilot is capable of executing these commands while processing the document. As a result, the worm can insert the same malicious payload into newly created or edited documents, facilitating its spread during file sharing. This manipulation of generative AI systems does not depend on traditional exploitable vulnerabilities or execution of malicious code, rendering it a novel method of attack.

The researchers clarified that while this demonstration serves as a proof of concept, no evidence suggests real-world implementation of this method in active attacks. However, its implications for security in AI-assisted document creation emphasize the need for vigilance and proactive countermeasures.

Other Stories From This Week

With reporting from ISMG’s Emilia David in Manhattan.