News items I have found interesting and informative.
OpenAI acknowledges its next-generation models may reach zero-day and intrusion-level capabilities, NIST moves to define a threat and mitigation taxonomy for AI agents, Tenable shows how a simple prompt injection against Microsoft Copilot Studio led to data leakage and fraud, and the Cloud Security Alliance publishes detailed guidance on AI prompt guardrails for enterprise GenAI.
Google patches a Gemini Enterprise flaw that could expose corporate data, new research shows most LLM application risks evade traditional code scanners, Cisco launches an open-source A2A Scanner to secure AI agent networks, OWASP publishes a Top 10 for agentic applications, and Google adds a $20K bug bounty on top of its layered Chrome agent defenses.
NCSC warns prompt-injection flaws may be unfixable, 7AI raises $130M to scale AI-agent SOC tooling, and NVIDIA/Lakera release a unified safety framework for agentic AI systems.
In the last 24 hours, Google introduced layered defenses in Chrome to contain indirect prompt injection against its agentic AI features, the UK’s NCSC warned that LLMs will always be vulnerable to prompt injection, and new research revealed malicious VS Code extensions and AI-branded packages stealing developer data and credentials from high-value engineering environments.
Over the last 72 hours, AI security stories have centered on shadow AI forcing a rethink of governance, “IDEsaster” flaws in AI coding tools enabling data theft and RCE, FBI warnings about AI-generated virtual kidnapping scams, NATO’s use of an AI chatbot in large-scale cyber war games, and new evidence that AI image services and AI bots are reshaping data-leak and scraping risk.
New guidance on securing AI in OT from NSA/CISA, discovery of the Brickstorm virtualization backdoor used by state-linked actors, and industry warnings that detection must replace perimeter defenses as AI-driven threats accelerate.
I security for December 3, 2025 centers on new NSA/CISA guidance for safely integrating AI into OT, SandboxAQ’s launch of an AI security posture platform for shadow AI, research showing poetic prompts can jailbreak major models, Experian’s forecast naming AI as the top breach driver for 2026, and TÜV SÜD’s move to formalize AI penetration testing using NIST, OWASP, and MITRE ATLAS.
AI security on Dec 2 centers on a serious Codex CLI command-injection flaw, new data showing layered AI defenses still buckle under targeted attacks, Anthropic’s agents successfully exploiting real DeFi contracts, Android zero-days hitting AI-enabled mobile endpoints, and a major investment push into explainable AI-driven investigations for national security
Over the past five days, malicious and “dark” LLMs have lowered the bar for cybercrime, HashJack has exposed a new AI-browser injection vector, Olymp Loader continues to evolve as a stealthy MaaS platform, Anthropic’s Claude Opus 4.5 boosts agent capabilities amid ongoing safety concerns, and new CVEs and identity hardening moves remind defenders that AI security is inseparable from solid infrastructure and IAM hygiene.
“HashJack” exposes a new class of AI-browser prompt injection, malicious underground LLMs accelerate attacker automation, and a new fully-undetectable MaaS loader raises baseline threat levels.