A threat group known as JADEPUFFER made a dangerous leap: moving from hijacking traditional databases to directly targeting artificial intelligence infrastructure. Its new tool, dubbed ENCFORGE, is a Go-compiled ransomware that specifically encrypts AI model files: trained weights, vector indexes, training datasets, and other machine learning stack artifacts (formats such as .safetensors, .onnx, .gguf, and .pkl, among others).

The entry point was a critical vulnerability in Langflow (CVE-2025-3248), an AI workflow-building platform that allows unauthenticated code execution. The most striking detail, according to researchers at Sysdig, is that the attacker appears to have operated autonomously: in just five minutes and 24 seconds, it adjusted and rewrote several scripts until achieving a working attack path, evading real-time automated defenses.

The encryption combines AES-256-CTR with RSA-2048 key exchange, and the malware shows no data theft capability — its sole purpose appears to be destruction, not extortion through leaks.

Researchers at Pillar Security managed to "escape" the sandboxes of four of the most widely used AI coding assistants today, without attacking the sandbox directly. The trick: the AI agent writes a configuration file inside its controlled environment (something perfectly normal), but then a trusted external tool — such as Git, an editor extension, or a local daemon — reads and executes that file outside the sandbox, unaware it was tampered with.

One of the most serious flaws simultaneously affected Cursor, Codex, and Gemini CLI: all three could communicate with the Docker socket, a privileged service, allowing code execution directly on the host system. It has already been fixed. Google, on the other hand, rated two similar findings in Antigravity as lower severity, considering they require social engineering to exploit — while still acknowledging the technical quality of the research.

Hugging Face — used by more than 50,000 organizations to share AI models and datasets — confirmed it suffered a security breach carried out from start to finish by an autonomous artificial intelligence agent, without direct human intervention at each step of the attack.

The entry point was a malicious dataset that exploited two code-execution flaws in the platform's data-processing pipeline. From there, the attacker escalated privileges, stole cloud credentials, and moved laterally across several internal clusters over an entire weekend, carrying out thousands of automated actions spread across short-lived sandbox environments.

Curiously, Hugging Face itself also used AI to detect and analyze the attack, processing more than 17,000 logged events to reconstruct the incident timeline. So far there is no evidence that models, datasets, or public spaces used by users were altered.

The market for AI platforms aimed at Security Operations Centers (SOC) is growing faster than organizations' ability to properly evaluate them. According to figures cited by Prophet Security, between 80% and 95% of enterprise AI projects fail when moving from a controlled demo to a real production environment.

The core recommendation: before adopting an AI tool for your SOC, clearly define what you expect — a point solution, a new capability, or a different way of organizing security work? Demos tend to show clean alerts and accurate verdicts in seconds, but accuracy tends to degrade considerably under real production conditions.

Google DeepMind unveiled Gemini 3.5 Flash Cyber, a specialized, low-cost AI model designed specifically to find, validate, and fix software vulnerabilities at scale. It runs within CodeMender, Google's code-security agent, invoking the model many times in parallel to analyze far more code paths than a large, expensive model could cover.

In internal tests on Chrome's V8 JavaScript engine, the model found 55 confirmed unique vulnerabilities, outperforming both standard Gemini 3.5 Flash (47) and Anthropic's Claude Opus 4.6 (36). For now, access is limited to governments and trusted partners, given the dual-use risk of this kind of technology.