And as I start with security stories, OpenAI’s Atlas browser and similar AI agents are vulnerable to manipulation, according to research from cybersecurity firm SPLX. The study highlights how malicious actors can exploit these vulnerabilities by serving different content to AI crawlers, potentially launching smear campaigns or presenting false job qualifications. For instance, a test showed that a fictional candidate could receive inflated credentials when viewed by AI, allowing her to outperform other candidates despite having weaker qualifications. Additionally, a survey by the British Standards Institution revealed that only 17.5% of U.S. business leaders have an AI governance program in place, indicating a significant gap in responsible AI use compared to global counterparts.
Bleeping Computer highlights a significant security gap in Zero Trust architecture concerning AI agents, which are increasingly being used in organizations. These AI agents, capable of making decisions and accessing systems without direct human intervention, often operate under inherited credentials, lacking a clear identity or governance. According to Token Security, this creates risks as these agents may appear trusted but can act as security vulnerabilities, leading to potential breaches. The National Institute of Standards and Technology’s AI Risk Management Framework emphasizes the importance of identity in managing these risks. It suggests organizations must map, measure, manage, and govern AI agents with a focus on identity governance to mitigate threats effectively. Without a trusted identity, security teams struggle with accountability when breaches occur, making it essential for organizations to embed identity controls into every phase of AI deployment.
Ransomware payments have reached a historic low, with only 23% of breached companies agreeing to pay hackers’ demands. This decline continues a trend observed by Coveware over the past six years, where payment rates had previously peaked at 28% in early 2024 before dropping significantly. The decrease is attributed to organizations strengthening their defenses and increasing pressure from law enforcement not to pay ransoms. According to Coveware, over 76% of ransomware attacks in the third quarter of 2025 involved data exfiltration, marking a shift towards double extortion tactics. As larger enterprises bolster their security, threat actors may increasingly rely on social engineering techniques to gain access.
The United States has opted not to join over 70 countries in signing the UN Convention against Cybercrime, a significant treaty aimed at establishing a global framework to combat digital crime. According to UN Secretary-General António Guterres, cybercrime costs the global economy approximately $10.5 trillion annually, with the convention recognized as a crucial instrument to enhance collective defenses against such threats. The treaty addresses various issues, including terrorism and human trafficking, and introduces the first global standards for electronic evidence. However, the convention has faced criticism from tech companies and human rights groups, who argue it may infringe on privacy rights and enable authoritarian practices.
Why do we care?
AI agents are turning into a new kind of security problem. Researchers say OpenAI’s Atlas browser and others like it can be tricked—fed fake data or doctored résumés—and the AI believes it. At the same time, many of these agents inside organizations don’t even have a real identity; they borrow credentials and operate unchecked.
That’s a Zero Trust nightmare—because if your AI can act but not be tracked, who’s accountable when it screws up or gets hacked?
Add in a bit of good news—ransomware payments are way down—but don’t relax too much. Attackers are just changing tactics, stealing data instead of encrypting it. And while over 70 countries just signed a UN cybercrime treaty, the U.S. didn’t—so expect more global fragmentation on rules and enforcement.
Bottom line? MSPs need to get ahead of this. Build governance around AI agents, track their identities, and help customers treat them like users—not magic boxes. Because AI isn’t just helping businesses—it’s now part of the threat model.