6 posts
WithSecure identified GREYVIBE, a Russia-nexus threat group targeting Ukrainian entities using spear-phishing, ClickFix, and fraudulent websites. The group systematically leverages Generative AI to develop custom malware (PhantomRelay, LegionRelay, FallSpy) and obfuscators, blending state-aligned intelligence gathering with cybercrime ecosystem overlaps.
During March-April 2026, threat actors increasingly deployed commercial AI models for real-time offensive operations, including automated intelligence analysis, BEC drafting, and vulnerability exploitation. Key developments include the weaponization of agentic configuration files for persistent jailbreaks, the rise of AI-integrated PhaaS platforms like EvilTokens, and the mass harvesting of AI provider credentials. Furthermore, AI capabilities are compressing the vulnerability patch window, allowing attackers to weaponize newly disclosed CVEs within hours.
The article highlights the growing risk of prompt data leakage in Generative AI workflows, where sensitive information like PII, source code, and API keys are exposed through conversational interfaces. It outlines 12 common leakage scenarios and recommends a phased approach to implementing inline DLP, browser isolation, and content moderation to secure AI usage without hindering productivity.
Threat actors are increasingly leveraging Vercel's GenAI capabilities, specifically v0.dev, to rapidly generate and host highly convincing credential phishing pages. By combining AI-generated frontends with Telegram Bot API integrations for real-time credential exfiltration, attackers can deploy resilient, low-effort phishing infrastructure on legitimate cloud services that evades traditional detection mechanisms.
Unit 42 identified 18 high-risk browser extensions masquerading as GenAI productivity tools that function as remote access Trojans, infostealers, and spyware. These extensions exploit browser permissions to intercept API keys, exfiltrate DOM content, establish persistent WebSocket C2 channels, and dynamically route traffic via malicious proxy configurations.
Unit 42 researchers demonstrated that both open and closed-source LLMs remain vulnerable to prompt jailbreaking at scale using a genetic algorithm-based fuzzing technique. By systematically generating meaning-preserving variants of disallowed requests, researchers successfully bypassed content filters and model guardrails, highlighting the fragility of current AI safety mechanisms under automated adversarial variation.
