News items I have found interesting and informative.
New research and reporting highlight how indirect prompt injection and agent memory can create persistent insider-style risk, while coding agents remain susceptible to prompt-driven unsafe actions. Adjacent cybersecurity updates, including CISA KEV additions and critical flaws in self-hosted platforms, reinforce that classic exploitation paths still underpin most AI compromise scenarios.
Malicious browser extensions were reported stealing ChatGPT and DeepSeek chats, while new research explored data adulteration to reduce the value of stolen AI knowledge graphs. Adjacent risks include extension marketplace supply chain exposure in AI-powered IDEs and continued emphasis on resilience as attacker speed increases.
The FBI warns of rising AI-assisted fraud, Microsoft tightens guidance on secure token handling for AI agents, and adjacent cybersecurity trends highlight KEV remediation and post-holiday phishing risks that affect AI environments.
Security leaders emphasized that AI agents can behave like insider threats if not tightly governed, while prompt injection remains a durable risk for tool-using systems. The week’s outlook reporting reinforces that AI-driven phishing and faster exploitation cycles will pressure identity, supply chain, and resilience controls in 2026.
A new “Lies-in-the-Loop” attack pattern exposes risks in AI safety workflows, Ciphero raises funding for runtime AI security, Anubis launches an open-source AI firewall to block scraping, and Darktrace enhances defenses against AI-driven social engineering.
AI-generated decoy documents are being used in active espionage campaigns, Microsoft reports hundreds of systems compromised via React2Shell exploitation, and Tenable highlights ongoing gaps in prompt injection and AI data security controls.
Cisco warns of active exploitation of an AsyncOS zero-day, a study finds WAFs ineffective against modern React2Shell exploits, and SonicWall patches an exploited SMA1000 zero-day, all of which affect perimeter defenses that protect AI-related services and admin portals.
NVIDIA patches a critical remote-code flaw in Isaac Lab, CISA adds a Fortinet signature verification vulnerability to the KEV catalog amid exploitation, and Anthropic finds that just 250 malicious documents can poison LLM training.
New research suggests AI-generated phishing training can improve user detection performance, while a new on-prem AI security tool highlights the growing push to apply AI to investigations without cloud data exposure, bringing new integrity and governance considerations.
CrowdStrike unveils prompt-injection defense tooling, autonomous AI pentesting emerges, U.S. federal executive order centralizes AI regulation, and CISA updates foundational cybersecurity goals to strengthen resilience against AI-driven threats.