In Other News: Anthropic Breach Dispute, UEFI Flaw Hits Most Servers

Major infrastructure flaws, distributed artificial intelligence data leaks, and state-backed cyber intrusions surfaced across enterprise environments this week, highlighted by a critical firmware buffer overflow that compromises servers worldwide as detailed in SecurityWeek's report. The disclosures track a broad spectrum of digital and physical threats, including a sixteen-year-old flaw embedded in global boot firmware, cross-tenant cache contamination on distributed inference platforms, and unauthorized network interventions hitting defense facilities.

This briefing consolidates technical disclosures, exposed systems, and immediate defensive actions for systems administrators, enterprise security architects, and compliance leads. It examines systemic vulnerabilities in foundational server firmware, third-party software risks in artificial intelligence engineering teams, operational disruptions to federal election filings, and evolving automation frameworks weaponized by threat actors for network penetration and identity abuse.

What broke in In Other News

Firmware security suffered a severe breakdown with the disclosure of CVE-2026-97631, a critical sixteen-year-old buffer overflow vulnerability located in TianoCore EDK II. The open source code base underpins firmware distributions across major global suppliers including AMI, Insyde, and Phoenix. In cloud infrastructure, Cloudflare identified an inference failure within its Workers AI platform between August 27 and September 9. A configuration mismatch left the key-value cache of the GLM 5.3 model smaller than its advertised context window. Combined with an upstream defect in the SGLang inference engine, this flaw allowed requests longer than 681,280 tokens to read cached data from earlier sessions, resulting in 155 compromised responses containing third-party customer information.

Operational systems and municipal networks also experienced severe disruptions. Suspected Iran-linked threat actors breached refrigeration controls at approximately six United States military commissaries starting in late August, forcing a freezer at Fort Huachuca in Arizona into defrost mode to spoil provisions. Similar overrides previously disrupted 350 establishments in France. Simultaneously, the United States Federal Election Commission had to take its electronic filing system offline after detecting an intrusion attempt, blocking mandatory campaign expenditure reports ahead of the November 3 midterm elections until service was restored on October 28. In academic networks, the University of Copenhagen suffered an intrusion into a test environment on October 20 that exposed massive community records.

Who is exposed by In Other News

Enterprise data centers and server operators running network boot processes represent the largest exposed group. Because TianoCore EDK II code is embedded across hardware sold by international manufacturers, the flaw affects most enterprise servers worldwide, provided they use Preboot Execution Environment network booting. Organizations utilizing Cloudflare Workers AI for large-context processing on the GLM 5.3 model were similarly exposed to tenant data leakage. External identity risks extended to corporate human resources departments, highlighted by defense contractor employee John Allen Kurrle Jr. pleading guilty to wire fraud conspiracy after using his Top Secret security clearance and home address to help North Korean remote workers secure nine corporate technology roles.

Academic, energy, and government sectors also face widespread compromise. The University of Copenhagen confirmed that 525,000 current and former students and faculty members had personal information compromised, including names, contact numbers, email addresses, and national identity CPR numbers, with 20,000 personal identification numbers for facility access cards directly exposed. In France, energy provider ENGIE reported that external threat actors accessed customer marketing tools operated by sales partners, exposing French customer names, contract numbers, and energy supply point identifiers. Furthermore, enterprise websites monitoring automated traffic face pervasive exposure from a threat actor using 32 spoofed artificial intelligence crawler identities across 300 IP addresses, touching 48 percent of tracked organizations.

What to do now about In Other News

System administrators managing networked servers must immediately audit their boot configurations and prepare firmware updates. While the vulnerability was originally scheduled for coordinated disclosure in January, early code fixes published by AMI have accelerated public exposure, requiring server administrators to contact hardware original equipment manufacturers for emergency patches. If network boot capabilities are not strictly required on local subnets, organizations should disable PXE functionality inside the BIOS interface to neutralize the remote attack vector. Machine learning teams hosting their own serving stacks must update the SGLang inference framework to version 0.5.22 or higher to patch cross-tenant key-value cache disclosure vulnerabilities identified by security researcher Gadi Evron.

Enterprises must also address identity exposure, credential rot, and regulatory compliance requirements. The University of Copenhagen has reissued card PINs for 20,000 impacted individuals, and organizations seeing high bot traffic must deploy behavioral traffic analysis, as self-declared crawler user agents from OpenAI, Anthropic, Meta, and Perplexity can no longer be trusted. Machine builders preparing for the European Union Cyber Resilience Act and Machinery Regulation can deploy Rockwell Automation SecureOT platform enhancements, which provide asset-level component inventories, automated vulnerability monitoring, and compliance mapping to the IEC 62443 industrial control standard. Security teams using command-line assistants should also inspect local environments to block unauthorized grey-market proxy relays altering official CLI binaries.

How the In Other News issue works

The TianoCore EDK II vulnerability functions as a classic memory corruption flaw triggered over the local network. An attacker on the same network segment transmits an oversized ICMP error packet to a server executing a PXE network boot sequence. The system firmware fails to properly bound the incoming message length, causing a buffer overflow in memory before the operating system initializes. If the adversary possesses knowledge of the target system's precise firmware build, they can execute arbitrary code with maximum processor privileges, bypassing Secure Boot and hardware-level protections. In the Cloudflare AI defect, memory allocated for the inference engine failed to match token limits, allowing unbounded queries to read unreleased attention cache buffers.

Advanced threat groups have increasingly integrated automated tools to execute lateral movement and persistence. Sophos revealed that the Gold Ogre ransomware cartel operating Aur0ra malware leveraged AI agents within Cursor to perform automated network discovery, privilege escalation, and lateral traversal, having bypassed model safety guards by masquerading as an authorized red team. Cisco Talos documented the NIK C2 framework, an artificial intelligence command-and-control platform deploying Windows implants that intercept keystrokes, exfiltrate stored credentials, and trigger screen-locking malware. In parallel, threat actors established grey-market proxy networks across mainland China that install authentic Codex and Claude Code command-line interfaces while rewriting endpoint network configurations to bypass sandbox controls and stream data externally.

What is still unknown about In Other News

Significant uncertainty remains surrounding the precise intrusion blast radius claimed by startup Cathedral against Anthropic. Cathedral, founded by former DOGE personnel, reported that its autonomous agents penetrated Anthropic internal networks in under 12 hours via third-party enterprise tools, allegedly accessing employee salaries, meeting notes, and Slack conversations regarding advanced frontier models. Anthropic disputed this framing, stating that the intrusion was isolated to a small subset of IT, sales, and human resources applications within a single day, maintaining that core model weights, training source code, and Claude user chats were never accessed. Anthropic also accused the startup of violating bug bounty rules by withholding operational records.

Attribution and ultimate casualty totals also remain unconfirmed across several international incidents. The University of Copenhagen has not identified the criminal operators or motives behind its October 20 breach, and French energy utility ENGIE has withheld the total number of customer accounts compromised through its marketing partner. United States Cyber Command and Department of Defense law enforcement units continue joint investigations with French authorities to determine whether state-sponsored Iranian actors coordinated the commissary freezer shutdowns. Finally, data center operators await firmware release schedules from Insyde, Phoenix, and major server vendors, whose delayed patch cycles leave hardware exposed following the disrupted January disclosure timeline.