UEFI Flaw Exposes Most Servers to Pre-Boot Attacks, Cloudflare AI Data Leak Reported

A critical, 16-year-old buffer overflow vulnerability (CVE-2026-97631) has been discovered in TianoCore EDK II, the open-source firmware code used by the vast majority of server vendors and data centers worldwide. This flaw, which impacts systems through firmware suppliers like AMI, Insyde, and Phoenix, could allow attackers with local network access to execute code at the highest privilege level before the operating system even loads.

The vulnerability is exploitable during the network boot process (PXE) by sending an oversized ICMP error message. If an attacker knows the target's firmware build, they can potentially bypass security measures such as Secure Boot. While only machines utilizing network boot are exposed, the flaw's widespread use in server infrastructure means a significant portion of the global server market is at risk. The vulnerability was discovered by researchers from Anthropic and Google, and its public disclosure was accelerated after AMI released a code fix, prompting organizations to seek firmware updates from their respective vendors.

In a separate incident, Cloudflare has reported a data leak affecting its Workers AI platform. Between August 27 and September 9, 2026, 155 responses from the GLM 5.3 model inadvertently contained data from other customers' recent requests. This occurred when prompts exceeded the model's KV cache size due to a configuration mismatch, combined with a bug in the SGLang inference engine that allowed over-length requests to access residual cache data from previous sessions.

The leak was specifically tied to prompts exceeding 681,280 tokens. Cloudflare has since remediated the issue by addressing the configuration mismatch and the SGLang inference engine bug, ensuring better tenant isolation and preventing future cross-customer data exposure. The company has notified affected customers and emphasized that the incident was limited to specific, long-form prompts on the GLM 5.3 model.

Additionally, the article touches upon other security developments, including suspected Iranian state-sponsored hacking targeting U.S. military freezer systems, an attempted intrusion at the U.S. Federal Election Commission that briefly took its systems offline, and a data breach at the University of Copenhagen exposing personal data of over half a million individuals. These diverse incidents underscore the persistent and evolving nature of cyber threats across various sectors.

Another notable mention is the discovery of fake ChatGPT and Claude bots actively scanning websites for sensitive information, such as configuration files and cloud credentials. Cequence Security observed a single threat actor using numerous AI bot identities to probe for secrets, highlighting the emerging threat of AI being weaponized for reconnaissance and data exfiltration. The ease with which these bots can impersonate legitimate AI services poses a significant challenge for detection and defense.

Finally, the report briefly notes a dispute between cybersecurity startup Cathedral and AI company Anthropic regarding the scope of an AI-led intrusion. Cathedral claims its AI agents accessed Anthropic's internal systems, while Anthropic states the access was limited and that core AI data remained secure, also accusing Cathedral of violating bug bounty program rules. This highlights the complex interactions and potential conflicts arising from the rapid development and deployment of AI in cybersecurity.