A fundamental shift: How AI cyber threats are reshaping cybersecurity programs

How AI cyberthreats are reshaping cybersecurity operating models

Artificial intelligence is disrupting how cybersecurity operating models and programs are built. New evaluations and incidents sharpen the central question for leaders: Can security and governance programs keep pace with AI-powered threats? As organizations accelerate AI adoption, threat actors can use the same capabilities to operate with greater speed, scale and sophistication. CEOs, chief information officers, chief information security officers and boards should understand what is shifting, engage in the decisions that follow and support the agility their organizations need to keep pace.

The answer is not to rebuild the cyber program around every model release. It is to prepare for the class of capability now emerging: rapid vulnerability discovery, persistent probing and increasingly autonomous actions. Organizations should also ask whether their defences are being tested by AI-enabled professionals who can evaluate whether the business can withstand attackers operating at machine speed.

How quickly are AI cyber capabilities advancing?

Recent research moves the discussion beyond one-shot vulnerability discovery. In April 2026, the U.K. AI Security Institute reported that Claude Mythos Preview completed expert-level capture-the-flag tasks 73% of the time. It also became the first model to complete the institute’s 32-step simulated corporate network attack, succeeding in three of 10 attempts. The institute cautioned that its ranges lacked active defenders and some protections found in real environments, but concluded that the model could autonomously attack small, weakly defended and vulnerable enterprise systems after gaining network access.

That finding does not mean every frontier model can compromise a well-defended enterprise. It means the boundary between research capability and operational cyber capability is moving quickly. Work that once required experienced professionals and sustained efforts can increasingly be compressed into automated or semiautonomous workflows. The result is a larger population of attackers that may be able to apply sophisticated techniques.

What do recent AI security incidents reveal?

Recent incidents show that AI security risk includes not only how threat actors use models, but also how capable agents behave when they receive tools, credentials, network access and incomplete boundaries. These events occurred in evaluation or testing contexts, not conventional malicious activities. Even so, they reveal control failures that enterprise AI programs should anticipate.

OpenAI reported that, during internal cybersecurity evaluations in July 2026, research models circumvented controls intended to isolate them from the internet, exploited shared infrastructure and compromised parts of OpenAI and Hugging Face infrastructure. OpenAI described the incident as evidence that highly capable agents can work around technical controls and take actions no human directed when safeguards are insufficient. The company responded with more isolated sandboxes, tighter internet and model-weight access, stronger alignment requirements and greater investment in monitoring.

A separate cybersecurity evaluation involving Google Gemini, reported by Reuters, showed a related control problem. Gemini accessed systems at three companies outside the intended test scope after guessing credentials in one instance and locating publicly exposed credentials in two others. Google said the model stopped its activity in all three cases, affected entities were notified and testing processes were changed. Nonetheless, the lesson is larger than one model or test: scope enforcement, credential controls and human oversight matter whenever an AI agent can act on connected systems.

The emerging risk is not simply AI as an attacker. It is AI connected to tools, identities and infrastructure without controls that reliably enforce scope, privilege and accountability.

How AI cyberthreats change security operations

Mature programs have always assumed fast, capable adversaries. What is changing is the number of actors operating at that level. AI can help attackers identify complex vulnerabilities, turn findings into working exploits and sustain high-volume activity. Incidents may no longer come in discrete waves; they may come continuously. Triage built for human-paced queues can break under continuous, machine-speed pressure.

Recent research suggests the window between vulnerability discovery and exploitation is shrinking rapidly. As frontier AI models become more capable of identifying weaknesses, generating exploits and supporting multistep cyber operations, organizations should assume they have less time to detect, prioritize and respond. The challenge is no longer simply keeping pace with technology change. It is governing critical attack surfaces, identities and response processes at the attacker’s pace.

Mean time-to-exploit has fallen from 2.3 years in 2018 to roughly 56 days in 2024 and is expected to drop below 24 hours by 2028.

The same technology can help defenders shorten vulnerability discovery, threat detection and response. The advantages are not equal, however. Attackers do not have to manage change-control windows, regression risk, downtime limits or enterprise patching at scale. Security teams do. Organizations should expect vulnerability disclosure, exploit development and automated attack attempts to almost converge rather than occur in sequence.

Where current cybersecurity programs fall short

Most programs were designed around slower exploitation and periodic testing. Five familiar gaps become more consequential when AI compresses the attack cycle:

The same root cause sits underneath all five patterns: security operations function as a set of activities rather than as a continuously tested discipline. Controls that are not measured, challenged and improved continuously and at the attacker’s pace can create confidence without evidence. In an environment where AI compresses the attack cycle, resilience depends less on adding controls and more on proving that existing defences perform as intended.

How should organizations prove AI cyber readiness?

AI cyber readiness is the ability to detect, contain and recover from AI-enabled attacks while governing AI systems that can act across data, identities, tools and infrastructure. Readiness is proven through realistic testing, measured response and clear accountability. Traditional penetration testing and periodic assessments remain valuable, but organizations should also test how defences perform against AI-assisted attacks such as credential misuse, persistence or sustained operational pressure. Red teaming and similar approaches have proven effective at advancing even well-established security programs.

A practical baseline includes a tested same-day containment scenario for each crown-jewel asset, named ownership for continuous vulnerability operations, and validated time-to-detect and time-to-contain measures. Leaders should request trend lines and scenario results, not snapshots of tools deployed or policies reviewed.

Programs should prioritize egress filtering, deep segmentation, reliable asset management, phishing-resistant multifactor authentication (MFA), identity governance and privileged access management. These controls raise attacker cost and limit the blast radius when an identity, agent or system behaves outside its intended scope. AI-enabled offensive testing should validate the maturity of your program through red teaming, purple teaming and assume-compromise exercises.

The objective is not simply to accelerate individual security activities. It is to govern identities, privileged access, internet-facing assets, response processes and AI-enabled systems at the pace an attacker can act.

Why AI governance is now a cybersecurity requirement

AI governance is the system of policies, accountability, risk classification and oversight used to direct responsible AI adoption. As AI systems gain access to enterprise data, identities, tools and infrastructure, governance becomes inseparable from cybersecurity. A policy that defines acceptable use but does not address permissions, monitoring, incident response and testing leaves operational risks unresolved.

Organizations should connect AI governance with cybersecurity, privacy, identity management, software development and third-party risk. That connection helps leaders answer four practical questions about both internally developed AI and capabilities provided by vendors:

• Which systems may act autonomously, and what data and tools may they access?
• How are scope, identity and privilege enforced?
• Which actions require human approval, and who can stop an agent?
• How are AI activity and security events monitored, including when models or capabilities are provided by third parties?

Third-party oversight is especially important as AI capabilities are embedded in software that organizations already use. Before enabling those features, leaders should understand what data the vendor can access, how that data and the underlying models are protected, and whether the vendor’s security, development and data management practices align with the organization’s requirements. Oversight should continue as vendors update their models, integrations and use of organizational data.

The OpenAI Hugging Face and Google Gemini incidents also underscore the need to govern evaluation environments. Reduced safeguards, permissive connectivity, exposed credentials or ambiguous testing scope can turn a legitimate exercise into unauthorized activity. Testing should use isolated environments, explicit authorization, least privilege, monitored egress, kill switches and professional oversight proportionate to the model’s capabilities.

An AI-ready operating model: Five connected motions

The AI-ready technical security program rests on five connected motions. Each produces evidence. Each sharpens the next.

Fight AI with AI. Use AI-assisted testing across code, pipelines and infrastructure, then pair it with AI-enabled defence that accelerates detection and containment. Stress the program with AI-enabled adversarial simulation, red teaming and assume-compromise exercises that reflect how a modern attacker works.

Harden and test the basics. Consistently apply egress controls, segmentation, asset management, identity governance and phishing-resistant multifactor authentication (MFA). Validate them with AI-accelerated adversarial simulation so every decision is tied to a tested outcome rather than a control catalog.

Recalibrate defence technology, risks and triage metrics. Measure time-to-detect, time-to-contain and triage throughput against scenarios that assume simultaneous activity and potential same-day exploitation. Risk scoring and patch service-level agreements (SLAs) may need to be reassessed so detection coverage and response capacity align with the new reality.

Stand up vulnerability operations (VulnOps). VulnOps treats asset discovery, identification, prioritization and remediation as a continuous operating function rather than a periodic project. This operating discipline helps organizations respond when exploit windows shrink from months to days or hours.

For middle market organizations, the priority is disciplined execution: continuous asset visibility, risk-based prioritization, timely remediation and targeted AI-enabled testing of critical systems. For enterprises, the same operating model scales across business units, regions and trust boundaries, with combined metrics that defend investment at the board level with evidence rather than narrative. In both cases, the objective is the same: reduce exposure faster than adversaries can exploit it.

Rewire secure software development. Technical debt is now an attack surface. Embed threat modelling, code review, provenance and test gates into the development lifecycle, apply security controls across machine learning operations, limit agent permissions and use zero trust segmentation to contain the impact of compromise.

What the AI era requires from modern cybersecurity programs

AI-accelerated attacks require organizations to rethink the assumptions behind their cybersecurity operating models. Recent research and real-world incidents suggest that security controls must perform under machine-speed pressure, identities and access paths must be constrained, and AI systems must operate within governance structures that enforce accountability and limits.

Organizations do not need to predict every future AI capability. They need an operating model that can absorb new capabilities without requiring a strategy rewrite. That means continuously validating controls, reducing viable attack paths, strengthening identity and vulnerability operations, and selectively applying AI where it meaningfully improves speed, visibility and decision-making.

The organizations most likely to succeed in the AI era will be those that treat cybersecurity as a continuously tested operational discipline rather than a collection of controls, projects and point solutions.

How RSM can help organizations keep pace

RSM helps organizations identify where AI-accelerated attack paths already exist; assess whether detection, response, identity and vulnerability operations can function at modern attacker speeds; and design the operating motions needed to close the gap.

Our work spans four connected priorities:

• Test. Our human-led, AI-accelerated adversarial simulations help identify attack paths, expose security blind spots and validate whether defences perform as intended against modern threats.
• Defend. RSM Defense uses AI-enabled capabilities to accelerate threat investigation and response while maintaining professional oversight, judgment and validation throughout the process.
• Transform. Our AI-enabled identity and cyber transformation services help organizations identify risky identities, excessive privileges, unintended access paths and configuration weaknesses that can increase cyber risk.
• Govern. We help establish AI governance and risk management frameworks that provide the policies, accountability, risk classification and oversight needed for responsible AI adoption.

Effective AI security risk management requires more than technical controls alone. It requires coordination across governance, security, identity, operations and response functions. RSM’s integrated capabilities are designed to help organizations address these interconnected challenges and strengthen resilience in an increasingly AI-driven threat landscape.