Current reading on AI security, from government agencies, standards bodies and the security press. Reviewed 6 November 2026.
A short government blog on why employees reach for unapproved AI tools and why understanding that is the first step to managing it. Useful framing for any organization writing its first AI use policy.
Verizon's annual breach report press release, which also reports that staff use of unapproved AI tools rose sharply and is now a common source of data loss. Gives leadership a credible number to bring to a budget or risk discussion.
Joint government guidance on giving AI agents access to systems and data, starting small and keeping human oversight. Read it before connecting an agent to company mail or files.
Explains how to limit what an autonomous AI system can do and how to watch it, after recent incidents of agents acting outside their brief. Plain enough for executives and boards.
Shows how a poisoned tool description can steer an agent into a harmful action, and lists controls such as approved tool sources and human sign-off. A concrete picture of connector risk.
Google looked at public web pages for hidden instructions aimed at AI tools and reports what attackers are actually trying today. Helps separate real risk from hype when employees use AI to browse and summarize.
Trade press report on a researcher-found flaw, since fixed, showing that Copilot can reach whatever the signed-in user can reach. A good reason to tidy permissions before a rollout.
Describes flaws in the consumer version of Copilot that could pull data from connected apps. Relevant to any organization where employees sign in to personal AI accounts on work machines.
Microsoft's own checklist for finding and fixing oversharing in SharePoint and Teams before Copilot goes live. The practical counterpart to the two news items above.
News coverage of a Pindrop survey of security leaders on deepfake attacks and their cost. The survey is vendor research; read it for the scale of the problem and the basic advice on training and verification.
The FBI's yearly crime figures, with its first section on AI used in crime, including voice cloning to request wire payments. Official numbers for any briefing on payment fraud.
A summary of the final changes to the EU AI Act, including the move of high-risk deadlines to December 2027 and August 2028. Useful for any organization with EU customers, operations or staff.
Reports on a Grant Thornton survey in which governance and compliance barriers were the most cited reason AI falls short. Useful in any sector for justifying time spent on governance.
Walks through the building blocks of an AI policy: scope, approved tools, data rules and oversight.
A downloadable starting policy to adapt. This is a template, not an article.
These links go to third-party sites. Agentiv Labs is not affiliated with their publishers.