Agents of chaos

The dream of self-directed AI is becoming real, and companies can’t wait to turn their agents loose on everyday workflows. Standards like the Model Context Protocol are helping agents move from chat to action, while new frameworks promise smoother integrations and smarter automation. But as these agents grow more autonomous—and start teaching themselves—so do the risks. From compliance to cybersecurity, organizations must now learn how to govern what they’ve unleashed. This month’s stories explore both sides of that frontier: the breakthroughs and the breakdowns of agentic AI.

Top picks for generative AI readers on InfoWorld

How self-learning AI agents will reshape operational workflowsCompanies are salivating at the idea of letting autonomous AI agents off the leash, and their emerging ability to learn from their own experiences (and mistakes) can make that vision a reality.

How MCP is making AI agents actually do things in the real worldThe Model Context Protocol (MCP) is an emerging standard that allows generative AI systems to interact with other services and data stores, and it can help release AI agents from their walled garden into the real world of productive work.

How to write nonfunctional requirements for AI agentsIf AI agents are to succeed, we need to develop them with their users in mind, just like any other type of computer program. Isaac Sacolick explores how we can build nonfunctional requirements focusing on areas of performance, compliance, security, and observability into the development process.

‘Blame the intern’ is not an agentic AI security strategyWe’ve already seen real-world instances of AI agents giving themselves elevated permissions and running amok. Management can’t simply shrug and blame the agents: they need to build a safe environment in which those agents operate.

GenAI news bites

Microsoft unveils framework for building agentic AI apps

Claude Sonnet 4.5 coding model improves agentic capabilities

Microsoft Marketplace opens for AI apps, agents

Google releases MCP server to Data Commons public data sets

More good reads and generative AI updates elsewhere

Google DeepMind launches an AI agent to fix code vulnerabilities automaticallyCodeMender uses Gemini Deep Think models to create an autonomous agent capable of debugging and fixing complex security flaws, and has already submitted patches to open source projects.

Chatbots play with your emotions to avoid saying goodbyeEven in the age of AI, “time on site” is the metric most internet services value—so it shouldn’t be surprising that chatbots will pull out all stops to keep you from shutting them down.

Recruiters use AI to scan resumes. Applicants are trying to trick it.As recruiters and HR departments make AI an integral part of their recruiting process, job seekers are trying to hack their way to the top of the pile with hidden prompts embedded in their resumes.