Cato AI Labs disclosed two critical RCE flaws in Cursor IDE (CVE-2026-50548, CVE-2026-50549, CVSS 9.8) collectively named DuneSlide that allow attackers to break sandbox isolation via prompt injection, affecting a widely-adopted AI code editor used by Fortune 500 developers. Developers using Cursor for AI-assisted coding face zero-click remote code execution risk; sandboxed execution is no longer a reliable security boundary.

DeepKeep disclosed InkJect, a visual prompt injection class that hides malicious instructions in image metadata to bypass model guardrails across leading multimodal systems. Teams deploying vision-language models must now treat image inputs as untrusted user-controlled code, not safety-filtered text alone.

A fully functional ransomware sample (deepseek_python_20260125_da0631.py) uploaded to VirusTotal demonstrates AI-generated malware exploiting Chromium APIs across desktop and mobile platforms. The ease of generating functional cross-platform malware via LLMs signals a shift in attacker tooling maturity and payload accessibility.

Academic research confirmed machine learning models cannot reliably distinguish authorized from unauthorized input, leaving them persistently vulnerable to prompt injection attacks at scale. Defenders face a fundamental limitation: no model-native solution exists to filter adversarial prompts reliably.

A 2026 survey ranked self-mutating malware (55.9%), public LLM data leakage (53.5%), and AI-driven evasion techniques (52.5%) as high or extreme risks, yet organizations lack detection and mitigation playbooks. Enterprise security teams must rebuild incident response workflows to handle AI-assisted attacks and data exfiltration at LLM inference scale.