Agents Unbound: AI’s Latest Moves Demand a New Kind of Leadership

OpenAI’s announcement that its newest model, GPT-5.6, will finally launch following a US government security review highlights how advanced AI is now treated as a national security matter ([1]). Officials pressed OpenAI to delay GPT-5.6’s release amid concerns about the model’s potential misuse, marking the first time regulators intervened preemptively in an AI deployment ([2]). This unprecedented vetting process signals that public sector oversight is becoming a factor in how and when top-tier AI systems reach the market.

Meanwhile, other tech players are racing to debut their own agentic AI. Elon Musk’s newly renamed company, SpaceXAI (formerly xAI), has confirmed it will soon release Grok 4.5 – a model billed as “Opus-class,” offering reasoning abilities on par with the best incumbents but with significantly faster responses ([3]). Grok 4.5 has been refined using massive feedback from Tesla and SpaceX operations, with a core focus on token-processing efficiency to attract enterprises sensitive to latency and cloud costs ([4]). This move adds to a growing roster of frontier models, ensuring that competition at the cutting edge of AI capabilities remains fierce.

Even Google is quietly boosting the autonomy of its AI. This week Google DeepMind added background task execution, remote tool support, and custom functions to its upcoming Gemini AI’s managed agent platform ([5]). These upgrades make it easier for Google’s agents to carry out multi-step operations and integrate with external applications without human intervention. Taken together, the message for leaders is clear: ever more powerful and efficient AI agents are arriving, and they will be subject to greater outside scrutiny as they do.

Major enterprise software platforms are rapidly integrating AI agents into daily workflows. In the past 48 hours, Slack rolled out a powerful update connecting its Slackbot with the entire Salesforce software suite ([1]). This enables employees to use natural language commands in Slack to execute tasks that span multiple systems – for example, pulling up live customer data from Salesforce CRM, generating a sales report in Tableau, or even initiating a DocuSign contract – all without leaving their chat window ([2]). By acting as a conversational orchestrator for analytics, CRM, and collaboration, Slack’s AI assistant is collapsing cross-departmental workflows into a single interface, which could dramatically speed up response times and break down silos.

Microsoft has taken a similar leap by embedding agentic AI into the tools that sales and service teams use every day. The company’s new “Sales Agent” and “Service Agent” are now generally available inside Dynamics 365 and Microsoft 365 Copilot, where they are designed to enhance customer interactions and streamline routine processes in-context ([3]). Rather than introduce a standalone bot, Microsoft’s approach weaves AI assistance into the existing email, calendaring, and CRM interfaces that employees already know. The goal is to allow sales reps and support staff to work smarter – for instance, by automatically summarizing customer inquiries or drafting responses – without any disruption to their normal workflows.

AI "coworkers" are also becoming more ubiquitous tools for knowledge workers. This week, Anthropic expanded its Claude Cowork assistant from desktops to web and mobile platforms, enabling the agent to persist across devices and perform multi-step tasks in the background even if a user goes offline ([4]). Professionals can now initiate a complex project on their laptop – such as drafting a client presentation or analyzing a lengthy contract – and then monitor or refine it from their phone as the AI continues working autonomously on their behalf ([5]). According to Anthropic, the vast majority of users are leveraging Claude Cowork for everyday analytical and writing tasks rather than just for coding assistance ([6]). This trend suggests that AI agents are fast becoming always-on team members, handling administrative and creative duties so human employees can focus on higher-value work.

Retail is emerging as a key proving ground for autonomous agents. On July 7, Salesforce officially launched Agentforce Commerce, introducing dedicated “Shopper,” “Buyer,” and “Merchant” AI agents now generally available to its retail customers ([1]). These agents are deeply integrated with retailers’ own inventory, e-commerce, and order management systems. They do far more than answer questions – for example, a Shopper Agent can check product stock and store locations, provide personalized recommendations, and even initiate purchases or reorders for customers via chat, while a Merchant Agent can autonomously handle routine tasks like updating online catalogs or verifying shipping cut-off times ([2]). By connecting directly to real-time data and transactional systems, the agents can take action to drive sales and service outcomes, not just respond with information.

Evidence is mounting that such agent-led commerce can deliver significant benefits. During the 2025 holiday shopping season, AI assistants influenced roughly 20% of global online sales – about $262 billion in value ([3]). Retailers that deployed their own AI shopper agents saw sales grow 59% faster than those that did not, with AI-referred traffic converting at about 8× the rate of traditional social media marketing ([4]). These results underscore how quickly consumer behavior is shifting toward AI-mediated channels. Instead of ceding customer interactions to third-party AI platforms, forward-thinking retailers are embedding their own agents within popular chat and search interfaces – and on their own sites – to capture this new demand while maintaining control of the customer relationship and data.

As autonomous capabilities accelerate, the dark side of agentic AI is coming into view. In a stark example, cybersecurity researchers this week confirmed the first documented ransomware attack carried out entirely by an AI agent ([1]). Dubbed “JADEPUFFER,” the attack saw an AI system take advantage of an unpatched vulnerability in an AI workflow tool (Langflow) and then autonomously execute a multi-stage ransomware campaign – over 600 discrete actions – without any human participation beyond the initial deployment of the malware ([2]). The AI intruder moved at machine speed through reconnaissance, data theft, and encryption of critical databases, demonstrating an unprecedented ability to adapt and self-direct. Investigators noted that JADEPUFFER’s behavior exemplified a new class of "agentic" threat actor, where the entire intrusion lifecycle is handled by an AI – drastically lowering the cost and skill barriers for sophisticated cyberattacks ([3]).

Nation-state regulators are also raising flags about AI system vulnerabilities. A Chinese government agency this week warned that certain versions of Anthropic’s AI coding assistant (Claude Code) allegedly contain a hidden “security backdoor,” sparking concerns over the security of third-party AI tools ([4]). (Anthropic has disputed the claim, but the incident illustrates the level of geopolitical suspicion swirling around AI platforms.) Coupled with the U.S. government’s new interest in pre-approving advanced models, these developments highlight the critical importance of rigorous governance, validation, and oversight when deploying agentic AI – especially in high-stakes and regulated environments.

These rapid advances mean leaders must revisit how they manage technology and teams. Early movers like Cisco are already treating AI agent rollouts as enterprise-wide change programs rather than pure IT projects. By the end of this month, Cisco will have equipped 90,000 employees with a personal AI assistant – served via a mix of local and cloud-based models to address data security and cost concerns ([1]). Crucially, the company sees this as a test of culture and trust: technical capability alone won’t ensure value if employees don’t adopt the tools. Cisco is therefore investing in training, monitoring usage, and gathering feedback, effectively making the AI agent deployment an exercise in change management and workforce engagement, not just software installation ([2]).

Experts urge a similarly thoughtful approach from all organizations adopting AI agents. Simply giving employees powerful AI tools without redesigning workflows can backfire – for instance, by reducing human-to-human collaboration and institutional knowledge sharing ([3]). Instead, leaders should identify high-impact use cases where agents act as force-multipliers for human teams. The biggest efficiency gains are expected when AI tackles complex analytical, creative, or decision-support tasks – not just low-level administrative chores – thereby freeing up staff for more strategic work ([4]). Moreover, human judgment must remain “in the loop” for critical decisions. As Google veteran Addy Osmani advises, companies should let agents handle the repetitive "inner loop" execution, but keep humans accountable for the "outer loop" – setting goals, reviewing evidence, and making final calls on important matters ([5]). In short, to capture the upside of autonomous workflows, executives will need to pair technical innovation with process re-engineering, employee training, and robust risk management.