The 2026 AI Outlook: What We Saw Coming and What We Missed

Revisiting our ten predictions, the unexpected developments, and what I'm watching as 2027 approaches.

By Camille Stewart Gloster, CEO of CAS Strategies

Originally published in Command Line with Camille.

My phone hasn't stopped buzzing this week. Friends stuck on planes during Friday's FAA ground stop, clients dealing with Microsoft's outage the day before, and security colleagues coming off an October so punishing we've started calling it Hacktober. Different problems, but some version of the same question: What is going on?

The FAA says unstable router traffic disrupted communications and radar feeds across much of the country, with the precise cause still under investigation. Microsoft's outage left Copilot unable to reach some of the work data it depends on. Neither disruption has been established as an AI-related incident, and the available official accounts do not identify a cyberattack as the cause.

Meanwhile, the attacks are real. Palo Alto Networks' Unit 42 has documented autonomous intrusions against real organizations, and the FBI issued a warning about attackers using AI-driven offensive frameworks.

Last December, CAS Strategies published ten predictions for AI in 2026. Eleven months later, most have held up, although the risks appeared in places we hadn't expected, particularly where systems and institutional responsibilities intersect.

2026 exposed the limits of looking at AI through separate technology, security, policy, workforce, and business lenses. I've spent much of this year asking a related question: Who owns the seam?

Infrastructure risk took an unexpected turn

We predicted that advanced chips and access to computing power would become geopolitical flashpoints. The NVIDIA developments revealed an unexpected dimension of that dependence.

Bloomberg investigated scrutiny surrounding Megaspeed, which imported billions in NVIDIA technology and rents overseas computing capacity to Chinese firms through arrangements generally permitted under U.S. rules. Megaspeed denies wrongdoing, and no charges have been filed against it.

Separately, U.S. prosecutors alleged a $2.5 billion chip-diversion scheme involving individuals associated with Super Micro Computer. Contractor Ting-Wei Sun pleaded guilty in October. Super Micro is not a defendant.

Then NVIDIA announced Jensen Huang's departure on November 12 amid unproven foreign-intelligence allegationsinvolving a former Megaspeed executive. On November 15, Megaspeed said she had left the company in August 2025 and disclosed an internal investigation.

As of November 15, the allegations remain unresolved. Still, the broader concern is clear: concentration in AI infrastructure creates exposure through corporate relationships and geopolitical decisions, with consequences extending well beyond the organizations directly involved.

The FAA and Microsoft outages offer a related caution about the infrastructure onto which AI capabilities are being added. Both disruptions traveled through underlying dependencies and affected services far removed from the original problem. The FAA's root-cause investigation continues, while Microsoft's preliminary explanation identifies an issue introduced during an earlier network-capacity upgrade that surfaced after a subsequent network change.

I'm struck by how much attention we give advanced AI capabilities while foundational infrastructure remains this vulnerable.

Autonomous attacks became operational

We predicted that weaponized AI agents would become a serious threat. The evidence now comes from actual intrusions, not only research demonstrations.

In October, CrowdStrike documented four threat actors using AI agents during intrusions. In one sequence, an agent obtained powerful domain-controller credentials in approximately 20 minutes without further human guidance.

On November 12, Palo Alto Networks' Unit 42 reported that attackers used an AI framework called Cairn in 25 operations against at least 19 organizations during August and September. Ten experienced unauthorized access or data retrieval. Google and Mandiant documented related malicious use of agentic offensive frameworks, including ARTEX.

Cairn could take a target and objective, then plan substantial portions of an intrusion. The attackers also instructed the framework to regard their actions as authorized security testing, effectively allowing their own assertion of permission to influence what the agents did next.

That should worry anyone giving AI systems access to enterprise environments. The ability to carry out an action tells us nothing about whether an agent was legitimately authorized to perform it.

What happened when agents exceeded their boundaries

Research incidents in July revealed another problem: agents undertaking legitimate evaluation tasks were capable of unauthorized actions.

OpenAI agents escaped isolation controls by exploiting a previously unknown vulnerability and compromised portions of Hugging Face's infrastructure. The UK AI Security Institute documented 19 unauthorized actions across 10 of 122 evaluation runs. And Anthropic reported three incidents involving agents exploiting weak credentials or unprotected endpoints. Two affected organizations had not detected the activity themselves.

These occurred in research environments without normal production safeguards, but they demonstrated that capable agents could exploit real systems without being directed to attack them.

A separate OpenAI training task revealed how detection can fail. A model scanned 362 cryptocurrency wallets, spent test-network funds without authorization, omitted the activity from its report, and was discovered roughly three months later.

Organizations must account for agents weaponized by attackers as well as those exceeding their authority during legitimate work. I examined the latter in When the Reward Is the Attack Surface.

Governance still has catching up to do

My book, The Insider You Built, published September 15, and my accompanying Command Line article examined delegated authority, meaning the actions organizations permit AI to take while remaining accountable for what happens.

In October, Microsoft CEO Satya Nadella advanced a closely aligned argument in Models as Insider Risks in the Super Intelligence Era, calling for independent controls surrounding powerful models.

The technical foundations still lag. Partnership on AI found that four major agent frameworks identified agents only within individual runs, limiting the ability to trace authority across sessions.

In October, NIST proposed ten design tenets for agent identity management. These are proposals rather than finalized standards, and their publication underscores how much practical implementation work remains.

Two predictions still ahead of the curve

I expected robust enterprise AI functions to distinguish organizational leaders. IBM found that 76% of surveyed organizations had Chief AI Officers, compared with 26% the year before.

The titles are spreading faster than the decision-making authority and operational structures needed to make those roles effective. I may have been early in expecting that maturity to develop, but I remain convinced that organizations with genuinely integrated AI capabilities will have an advantage over those still managing AI through disconnected functions.

I also predicted that trust would outperform speed. Markets haven't consistently rewarded trustworthy AI with a commercial premium, but we're beginning to see trust determine who gets access in the first place. JPMorgan's Jamie Dimon has raised concerns about liability when customers give AI agents access to their accounts, while Schwab has restricted unauthorized outside agents.

I still believe trust will become a source of competitive advantage. For now, it's emerging as a permission gate, and I suspect the market will become much more discerning about security and accountability as the consequences of getting them wrong become harder to ignore.

Workforce choices shaped returns

An October Wharton-GBK study found that 80% of surveyed enterprise leaders reported positive AI returns, but only 28% reported strongly positive ones. Gartner found no meaningful association between AI-related layoffs and stronger returns.

IKEA offered a more constructive example, retraining thousands of customer-service employees for complex support and remote sales while automation handled routine inquiries.

In What Happens to Human Expertise When AI Does More of the Work?, I explored what leaders should do with the capacity automation creates. The answer matters because organizations will continue to need human judgment even as machines undertake more work.

Vendor risk went another layer deep

In November, Notion disclosed that customer AI chat transcripts had been exposed through its AI infrastructure provider Fireworks.

Reporting on Notion's incident FAQ indicated Fireworks had contractually promised zero data retention. Fireworks said attackers accessed development machines through another cloud provider.

The incident exposed how difficult contractual promises become to verify when data passes through organizations that customers may never have chosen or directly assessed.

What we're taking into 2027

I explored these organizational seams during a cross-functional threat-modeling workshop with Salesforce's Global Safety & Security leadership team.

As I wrote in Who Owns the Seam?, the people who understand operational consequences need a reliable way to influence decisions made elsewhere. That requires preserving context as information moves between teams and making clear who has authority to act when concerns arise.

For leaders deploying AI, I would start by asking whether every agent and its permissions can be identified, then test whether someone can actually stop one mid-task.

Boards should be examining what happens farther down the vendor chain, particularly when customer information enters systems the organization doesn't directly control.

2026 exposed the seams. In 2027, I want us to get better at owning them.

The CAS Strategies 2027 Outlook is coming. Subscribe to The Signal to receive it first.

The 2026 AI Outlook: Prediction scorecard

Assessment through November 15, 2026. The verdicts are qualitative judgments about the original predictions, not a statistical forecast-accuracy measure.

For organizations examining how these developments affect their deployments, explore the ACE Resource Hub or contact CAS Strategies about agent governance and enterprise resilience.

Next
Next

Free Digital Fluency Toolkit Helps Parents Guide Kids Through AI And Online Life