TODAY’S DISRUPTIVE BLOG

AI is collapsing the cost of cyberattacks. The disruption is not just smarter adversaries—it is the failure of security models that assume humans act at human speed.

Dennis G Perry, PhD, MBA | November 1, 2026

Strategic Thesis

The cybersecurity industry has built decades of defenses around a comforting assumption: successful attacks require scarce expertise, time, and operational discipline. AI is eroding all three barriers. Reporting published today describes an alleged amateur-led campaign against South Korean financial institutions that drew on multiple AI chatbots [1]. Reuters previously documented an attacker who used AI agents to compromise numerous websites at relatively low cost [2]. The lesson is not that every novice has become an elite hacker. It is that the marginal cost of discovering, adapting, and repeating attacks is falling—and organizations still rely on detection and approval processes designed for a slower threat environment.

Introduction

On November 1, The Washington Post reported that financial data from tens of thousands of South Koreans had been exposed following attacks affecting at least nine banks. Citing CrowdStrike’s investigation, the report described a suspect who allegedly used chatbots not just to aid the intrusion but also for subsequent illicit activities [1]. This reports an alleged campaign, not a judicial finding; the specific attribution and the degree of AI assistance remain subject to further corroboration.

More significant than any single case is the larger pattern. A tool that can explain a vulnerability, generate scripts, interpret errors, and suggest the next action can dramatically compress an attacker’s learning cycle. An experienced operator can use the same capabilities to increase throughput. AI does not eliminate the need for access or exploitable weaknesses. It makes the search for them cheaper and easier to repeat [1], [2].

Why It Matters Now

The attack surface is expanding just as the cost of testing it declines. Every public API, exposed administrative interface, stale identity, and unpatched workload becomes a candidate for automated examination. Even when a model makes mistakes, rapid iteration can compensate for many failures. An enterprise that measures patch cadence in weeks and approves exceptions in meetings is increasingly exposed to an adversary operating in minutes.

A second collision is happening inside the enterprise. SailPoint reported in October that 79% of surveyed organizations had AI agents in production while only 2% used identity-security tools purpose-built to govern them [3]. JumpCloud’s study found that among organizations using AI agents, 59% did not apply the same governance rules used for human users [4]. These are vendor-sponsored survey findings, not population-wide census results, but the direction is consistent: delegated machine authority is spreading faster than identity controls.

The Disruption

The real disruption is economic and architectural. Historically, defenders could impose costs through complexity: multiple systems, unfamiliar tools, separate credentials, and workflows requiring technical knowledge. AI reduces some of those transaction costs for attackers. At the same time, businesses are creating their own high-privilege agents and connecting them to the very systems criminals seek to exploit.

This creates two converging risks. Outside the perimeter, adversaries can industrialize reconnaissance and exploit adaptation. Inside it, legitimate AI agents may become an unintended execution channel if they have excessive permissions, accept untrusted instructions, or execute actions without independent policy verification. Salt Security’s October research reported that 49.8% of respondents confirmed or suspected an agent had taken an unexpected or unauthorized action in the prior year [5]. Confirmed and suspected incidents should not be conflated; the survey still identifies a governance problem worth testing.

The Fatal Assumption: Detection Equals Control

Security teams often equate visibility with safety. Logging every API call and displaying agent decisions can be useful, but an audit record written after a destructive command does not stop the command. A human-in-the-loop label is similarly weak if the human cannot meaningfully inspect each high-speed action or if approval is attached only to a broad task rather than the specific operation.

The alternative is independent enforcement before execution. For consequential actions, the system should evaluate verified identity, intended change, current state, allowed operating bounds, and policy authorization at the point of actuation. It should reject unsafe transitions even when the instruction originates from an authenticated controller or a trusted-looking AI agent. This is a design recommendation—not a claim that any one product has already solved the full problem.

Business Implications

First, technical debt becomes a more liquid liability: weaknesses that used to be hard to discover can be found and reused more economically. Second, privileged non-human identities become board-level risk, not merely an IAM configuration detail. Third, vendor claims that an agent is “secure” are insufficient without evidence of what it can do, who can revoke it, and what independently prevents prohibited changes. Finally, critical-infrastructure operators must distinguish loss of confidentiality from unauthorized physical or operational actuation; the latter may require hard safety interlocks and verified fallback states.

A 30-Day Executive Response

Week 1 — Inventory every production agent, service identity, external tool connection, and high-impact action path. Assign an accountable owner and identify agents that can write, delete, change permissions, transfer value, or affect physical operations.

Week 2 — Remove standing privilege where possible. Require short-lived credentials, scoped tokens, explicit allowlists, and separation between proposing an action and authorizing its execution. Test prompt injection and compromised-controller scenarios.

Week 3 — Implement pre-execution policy gates on the highest-consequence operations. Record signed authorization, relevant system state, command identity, and execution result. Define deny, hold, quarantine, and recovery behavior before an incident occurs.

Week 4 — Run an adversarial exercise in which an AI-assisted attacker and an overprivileged internal agent attempt the same unsafe outcome. Measure time to prevent action—not merely time to detect it. Report measurable residual exposure to executive leadership.

The Counterargument—and Its Limit

AI is also accelerating defenders: vulnerability discovery, code review, triage, and remediation all benefit. That is real progress. Yet symmetrical access to better tools does not produce symmetrical outcomes. Attackers need one workable path; defenders must maintain controls across every exposed pathway. The winning strategy is therefore not to prohibit AI or simply purchase more AI-powered detection. It is to reduce the amount of damage any successful instruction can cause.

Bottom Line

The next cybersecurity divide will not be between organizations that use AI and those that do not. It will be between organizations that let software exercise consequential authority on trust and those that require independent proof before action. Machine-speed threats demand machine-speed enforcement. Anything less is a faster way to document a failure.

References

[1] Gerrit De Vynck, “Anyone can be a super-hacker now: The world faces a cybersecurity reckoning,” The Washington Post, November 1, 2026. https://www.washingtonpost.com/technology/2026/11/01/wave-new-cyberattacks-is-just-getting-started/

[2] Raphael Satter, “How one hacker used AI agents and a few thousand dollars to break into scores of sites,” Reuters, October 19, 2026 (updated October 27). https://www.reuters.com/legal/government/how-one-hacker-used-ai-agents-few-thousand-dollars-break-into-scores-sites-2026-10-19/

[3] SailPoint, “SailPoint Report Finds 79% of Enterprises Run AI Agents in Production, Yet Only 2% Have Deployed Purpose-built Security,” October 6, 2026. https://investor.sailpoint.com/news-releases/news-release-details/sailpoint-report-finds-79-enterprises-run-ai-agents-production

[4] JumpCloud, “New JumpCloud Research Finds AI Access Is Outpacing Enterprise Governance,” October 7, 2026. https://jumpcloud.com/press/new-jumpcloud-research-finds-ai-access-is-outpacing-enterprise-governance

[5] Salt Security, “State of Agentic AI and API Security 2H 2026,” October 13, 2026. https://salt.security/press-releases/state-of-agentic-ai-and-api-security-2h-2026

[6] Gartner, “The AI Agents You Can’t See Can Still Create Risk,” September 30, 2026. https://www.gartner.com/en/articles/agentic-ai-oversight

Disruption is a Fact of Life

THE SUPER-HACKER IS NO LONGER AN EXPERT

Leave a Reply Cancel reply