Four CVEs of 8.3 to 9.9 on Pyodide sandboxes, ten vulnerabilities in llama.cpp, and a Zoom exploit chain built in less than 24 hours with public models.
The weekly review published on August 14, 2026 by eSecurity Planet aligns a series of incidents where AI is sometimes the target, sometimes the tool. The broadest architectural weakness affects Pyodide sandboxes: seven products that use them to isolate untrusted Python code have been found to be workable, generating four CVEs with CVSS scores ranging from 8.3 to 9.9, resulting in unauthorized access to the host environment.
On the local inference infrastructure side, ten vulnerabilities were disclosed in llama.cpp, including two llama-server flaws rated 9.2, related to memory safety flaws. At Black Hat 2026, researchers also presented critical flaws affecting the coding agents of Anthropic, Google and OpenAI, leading to remote code execution, credential theft, persistent prompt injection and supply chain compromises.
AI as an exploitation accelerator
Dubbed ZOOMSDAY, a chain of three vulnerabilities allowing remote code execution without interaction on Windows, macOS, iOS and Android was developed in less than twenty-four hours using public templates; Zoom has released patches. In the same dual-use vein, OpenAI introduced GPT-5.6-Cyber, a restricted model for exploit validation and defensive research, which successfully completed 95% of advanced cyber queries in testing — a rate that the company itself acknowledges as a risk.
Read the full article on the original source
Go to the source →