Security is the deployment bottleneck
CIOs now rank securing AI alongside malware, ransomware and phishing, and experts warn that autonomous agents break traditional security models unless run in sandboxed, least-privilege environments. A survey flagged AI security as a top priority and separate reporting shows three ransomware gangs accounted for 40% of attacks in March, reinforcing the concentration of high-impact threats. Media commentators recommend logging, approval gates, and strict runtime controls for agent deployments. (prnewswire.com) (digit.fyi) (podcasts.apple.com)
Security teams are discovering that the hard part of deploying artificial intelligence is no longer the model. It is the security wrapper around what the model is allowed to touch. (logicalis.com) Logicalis said on April 13 that 28% of chief information officers now rank artificial intelligence itself as a significant cyber risk, just behind malware at 33%, ransomware at 33%, and phishing at 30%. The company said its annual report surveyed more than 1,000 chief information officers worldwide. (logicalis.com) The same report said 77% of organizations suffered a cybersecurity incident in the past year, 57% said employees put data at risk through artificial intelligence use, and 34% said artificial intelligence created new security blind spots. Nearly half of respondents said they wished artificial intelligence had never been invented. (logicalis.com) An artificial intelligence “agent” is software that does jobs on its own, like reading files, calling tools, or changing systems without waiting for each next prompt. That autonomy is what changes the security problem, because the model is no longer just answering questions. (anthropic.com) The Open Worldwide Application Security Project said agents need the minimum tool access required for a task, with permissions scoped per tool and per resource. Its cheat sheet also warns that over-privileged agents can turn prompt injection into tool misuse, identity abuse, or code execution. (cheatsheetseries.owasp.org) NVIDIA’s artificial intelligence red team said the baseline controls for agentic workflows are network egress limits, blocking writes outside a defined workspace, and protecting configuration files even when a user tries to approve an unsafe action. Microsoft said on April 2 that it was open-sourcing an Agent Governance Toolkit for runtime policy enforcement, zero-trust identity, and execution sandboxing. (developer.nvidia.com) (opensource.microsoft.com) The ransomware backdrop is not getting calmer while companies sort this out. Check Point said 672 ransomware incidents were reported in March 2026, and three groups accounted for 40% of them. (infosecurity-magazine.com) Digit reported that Qilin alone was linked to 20% of March’s published attacks, while Akira and Clop helped make up the top three. Check Point counted 47 separate ransomware groups active that month, which means a fragmented market still produced a concentrated hit list. (digit.fyi) (blog.checkpoint.com) Standards bodies are moving in the same direction as vendors. The National Institute of Standards and Technology launched its Artificial Intelligence Agent Standards Initiative in March 2026 to push secure adoption and interoperability for agents that act on behalf of users. (nist.gov) The practical answer emerging across guidance is narrow and mechanical: run agents in sandboxes, give them least-privilege credentials, log every action, and force approval gates before high-risk steps. Until those controls are routine, security will keep deciding how far agent deployments can go. (cheatsheetseries.owasp.org) (developer.nvidia.com)
