The Director's Cut: Cyber Attacks Are Speeding Up. Boards Must Keep Up

AI agents, data-theft extortion, machine-speed attacks, and cross-sector crisis planning are reshaping cyber oversight. Here are four cyber & AI risk landscape developments and the questions directors should ask management now.

Rogue AI Agents Expose a New Control Problem

Recent incidents have raised new concerns about rogue AI agents. On September 18, Google confirmed that its Gemini model escaped a testing sandbox and breached the networks of three real companies during a capture-the-flag exercise. Between July and September, Google, OpenAI, Anthropic, and Meta all disclosed similar breakout incidents. The pattern suggests that as AI systems become more agentic, the central risk is no longer just what they can generate, but what they can do when connected to real tools, data, and networks.

For directors, this is a different governance problem from ordinary AI adoption. Agentic systems can chain actions across applications, data stores, and external services at machine speed. That raises a sharper oversight question: where is the organization allowing AI to act autonomously, under what permissions, and with what operational boundaries? A small failure in access design or approval logic can become a larger operational, data, or financial event.

Boards should expect management to treat AI agents as a containment and control issue before treating them as a productivity program. That means tightly scoped permissions, default-deny access, segmentation around sensitive systems, clear human approval thresholds, and rapid isolation if an agent behaves unexpectedly. Zscaler’s recent AI security assessments (here) reinforce the broader point that AI exposure is already ahead of governance in many environments. With agents, that gap becomes more consequential because the system can take action, not just produce content.

What Directors Should Ask Management:

• If an internal AI agent behaved unexpectedly, how would we detect it quickly and contain its access before it affected sensitive systems or data?

• How are we verifying that AI use across the organization is governed by effective access, segmentation, monitoring, and policy controls rather than informal adoption?

• If a third-party or internally deployed AI agent began chaining actions across systems, how quickly could we isolate it before it created operational, data, or financial harm?

The FBI Breach and the Rise of Data-Theft Extortion

On September 22, attackers breached FBIJobs.gov and stole an estimated 2–3 terabytes of data, later releasing a 5,000-record sample. Reporting indicates the data included records on current and former FBI employees and job applicants, including personal information. The attackers apparently exploited an Oracle PeopleSoft vulnerability in an FBI system.

Recent ransomware research from Zscaler shows how central data theft is to extortion. Data theft rose 276% year over year to nearly 900 terabytes, while the average ransomware payment climbed to just under $432,000. Extortion methods are also evolving quickly, and AI is helping attackers improve tooling, sharpen social engineering, and scale operations faster.

Directors should expect management to reduce unnecessary exposure, tightly control access to sensitive systems, and limit how far an attacker can move if one system is compromised. A Zero Trust approach can reduce attack surface, contain intrusions faster, and make large-scale data theft harder to execute.

What Directors Should Ask Management:

• How are we reducing the chance that attackers can steal enough sensitive data to create payment leverage even if systems remain operational?

Bitget Hack Shows How Quickly Control Failures Can Escalate

The $387 million hack of cryptocurrency exchange Bitget, attributed to North Korea-linked actors, was one of the year’s largest crypto thefts. After the breach, stolen assets moved rapidly across multiple blockchains, complicating recovery. The lesson extends well beyond crypto: once attackers gain access to privileged workflows, they can move at speed, exploit trusted infrastructure, and turn a single control failure into a major financial event.

Attackers increasingly operate at machine pace, and defenses built mainly around manual review will lag behind. Security operations will need more automation and AI-assisted detection, investigation, and containment to keep pace with attackers operating at machine speed. But human judgment remains critical. Management should be able to show that AI is accelerating response while high-consequence decisions, especially those affecting critical systems, customer data, or financial controls, remain under clear human oversight.

What Directors Should Ask Management:

• What decisions in our security response are automated today, and which high-consequence actions still require explicit human approval?

Top CEOs Treat AI Cyber Risk as a Crisis-Planning Problem

WSJ Pro Cybersecurity reports that a group of major U.S. companies is building cross-industry cyber crisis plans in response to AI-driven risk, reflecting a growing view that these threats now require CEO-level attention, not just technical oversight. The Alliance for Critical Infrastructure is expanding on the premise that as AI systems become more capable and interconnected, disruptions in one company or sector may spread more quickly across others.

No playbook will predict every way AI or other fast-moving technologies could trigger a crisis. Companies should still have a clear starting point: who needs to be involved, who has decision authority, how escalation works, and how coordination happens across business units and external partners. Boards should expect crisis planning that is adaptable, regularly tested, and built for incidents that can spread quickly across providers, sectors, and operational dependencies.

What Directors Should Ask Management:

• How are we working with peers, key providers, and relevant cross-sector partners to reduce the risk that a cyber incident escalates beyond our organization and disrupts the broader industry?

*****

Zscaler is a proud partner of NACD's Northern California chapter. We are here as a resource for directors to answer questions about cybersecurity or AI risks, and are happy to arrange dedicated board briefings. Please email rsloan[@]zscaler.com to learn more.