They Are Not Breaking In Anymore
- Rich Washburn
- 18 hours ago
- 8 min read
The CrowdStrike 2026 Global Threat Report and the end of the perimeter.
The CrowdStrike 2026 Global Threat Report is out, and the central message is this: attackers are not necessarily smashing through the front door anymore. They are logging in with valid credentials, moving through trusted cloud services, exploiting devices nobody is watching, and using AI to do all of it much faster.
CrowdStrike calls 2025 "the year of the evasive adversary." That is basically the whole report in one phrase.
The numbers worth remembering:
29 minutes — average time between initial compromise and lateral movement. That is 65% faster than the prior year.
27 seconds — fastest recorded breakout.
4 minutes — in one case, attackers began trying to exfiltrate data four minutes after gaining access.
82% — percentage of detections that involved no malware at all.
89% — increase in attacks attributed to adversaries using AI.
42% — increase in zero-days exploited before public disclosure.
266% — increase in cloud-focused intrusions by nation-state actors.
563% — increase in incidents using fake CAPTCHA lures.
141% — increase in spam email activity.
Read those again. The speed compression alone should reframe how you think about incident response. Twenty-nine minutes from initial access to lateral movement. Twenty-seven seconds in the fastest case. Four minutes to data exfiltration.
If your security operations center operates on a queue-and-triage model, that model is already obsolete. The attack finishes before the alert gets reviewed.
Identity is becoming the real perimeter
Attackers increasingly use stolen credentials, session tokens, legitimate SaaS integrations and approved administrative tools. Once they look like a valid user, traditional endpoint defenses have much less to grab onto. That 82% malware-free figure matters because it means security teams cannot rely on finding a malicious executable anymore. The "malware" may effectively be an employee account behaving slightly differently than usual.
This is the shift that matters most. For decades, security was built on a simple premise: something bad enters the environment, and we detect it. A file, a process, a network connection — something identifiable and foreign. The entire tooling industry was built around that assumption. But if 82% of detections involve no malware at all, the foreign object is no longer the primary threat vector. The threat is an identity. A credential. A session token. Something that already belongs inside the environment, being used by someone who should not have it.
We built defenses around the idea that an attacker would introduce something visibly malicious. The modern attacker does not need to. They need to look like business activity.
AI is an accelerant, not cyber sorcery
The report does not show autonomous super-hackers inventing entirely new forms of attack. It shows attackers using AI to accelerate reconnaissance, write and debug scripts, improve phishing and social engineering, create convincing fake identities, automate credential theft, and erase evidence and evade detection. That is still extremely consequential. AI is compressing work that once required skilled operators and significant time into something faster, cheaper and easier to scale. CrowdStrike measured an 89% increase in attacks involving AI-enabled adversaries.
I keep seeing this pattern everywhere. AI does not have to invent a fundamentally new capability to be transformative. It has to compress the time and cost of existing capabilities. That is what it is doing in cybersecurity, and that is what it is doing in every other field. The same logic I described in the one-person frontier lab applies here, just from the other side. One person with AI can now do what a hundred-person engineering team used to do. One attacker with AI can now do what a skilled intrusion team used to do. The leverage works in both directions. It amplifies competence and it amplifies malice. Good judgment gets leverage. Incompetence gets leverage. And so does intent to harm.
AI systems themselves are now targets
This may be the most strategically important part of the report. CrowdStrike says attackers abused legitimate generative-AI tools at more than 90 organizations, injecting prompts that caused those systems to produce commands used for credential or cryptocurrency theft. Attackers also exploited AI-development platforms, deployed malicious services impersonating trusted AI infrastructure and used compromised platforms to establish persistence or launch ransomware.
So the AI security problem is now two-sided. Attackers use AI against the enterprise, while also attacking the AI inside the enterprise. That includes models, agents, API keys, vector databases, notebooks, development pipelines, plugins and whatever permissions those systems inherit.
If you are building with AI — and at this point, who isn't — your AI infrastructure is now part of your attack surface. Every agent you deploy, every API key you provision, every model endpoint you expose, every vector database you spin up — that is all now part of the perimeter. And most organizations have not even begun to inventory it, let alone secure it. This is the same pattern as the cloud transition. Organizations adopted cloud infrastructure faster than they built cloud security. They adopted AI infrastructure faster than they built AI security. The gap between adoption and defense is where attackers live.
The cloud is not merely another endpoint
Cloud-focused intrusions rose 37% overall, and state-linked cloud intrusions rose 266%. Valid-account abuse represented 35% of the cloud incidents CrowdStrike observed.
Attackers understand that organizations still monitor laptops better than they monitor SaaS-to-SaaS connections, service accounts, cloud control planes, API tokens, identity federation, temporary infrastructure and AI-development environments. They are moving through the seams between these systems. The connections between services are less visible than the services themselves. The service accounts that hold persistent access are less monitored than the user accounts that trigger alerts when someone logs in from a new location.
The attacker does not need to defeat your cloud provider's security. They need to find the one integration, the one token, the one service account that was provisioned for a legitimate purpose and never revoked. Those objects are everywhere. Most organizations cannot even produce a complete inventory of them.
Edge devices are the soft underbelly
VPN appliances, firewalls, gateways and other internet-facing infrastructure are attractive because they often sit directly on the internet, run outdated software, produce weak telemetry, fall outside normal endpoint monitoring and provide immediate privileged access.
Forty percent of the vulnerabilities exploited by China-linked actors targeted edge devices, and 67% of the vulnerabilities they exploited provided immediate system access.
The device supposedly protecting the network may also be the least-visible computer in it. That sentence should make anyone running enterprise infrastructure uncomfortable. The security appliances — the things you bought specifically to protect the perimeter — are themselves becoming the entry point. And because they are security devices, they often sit outside the monitoring scope of the security team. You bought a firewall to protect the network. Nobody is watching the firewall.
Nation-state activity
China: CrowdStrike reports a 38% increase in China-linked activity across all sectors and an 85% increase against logistics organizations. The focus remains strategic intelligence collection, supply-chain access and persistent footholds in critical infrastructure and technology environments. Logistics matters because compromising it reveals how physical and digital systems connect — shipping, manufacturing, suppliers, ports, infrastructure, military dependencies and commercial flows. If you want to understand a nation's real capabilities, you map its logistics. If you want to undermine them, you compromise the systems that coordinate them.
North Korea: North Korea-linked incidents increased more than 130%. CrowdStrike highlights fake-worker and insider operations supported by AI-generated identities, along with financially motivated cryptocurrency theft.
One North Korea-linked operation stole approximately $1.46 billion, described in the report as the largest single financial theft ever reported. This is not ordinary cybercrime. It is effectively a state using hacking, fraudulent employment and digital-asset theft as a revenue operation. North Korea has figured out that the most efficient way to monetize a vulnerable digital economy is to combine AI-generated identities, remote work fraud and cryptocurrency theft into a single integrated operation. They are running a criminal enterprise at the scale of a national budget.
Supply-chain attacks are becoming normal infrastructure attacks
Attackers increasingly compromise software providers, public code repositories, development ecosystems, SaaS vendors and upstream service providers. That lets them inherit access to many downstream victims without attacking each one directly.
CrowdStrike treats supply-chain compromise as one of the defining tactics of the year.
This also means "we were not directly breached" is becoming a dangerously meaningless distinction. If your SaaS vendor was compromised, and your service integration still works, and the attacker is using that integration to move through your environment — you were breached. You just do not know it yet. The perimeter is not your network. The perimeter is everything you trust. And everything you trust is connected to everything else.
What companies should actually do
The practical response is not "buy more antivirus." Treat identity telemetry like endpoint telemetry. Monitor unusual authentication, token use, privilege changes, session behavior and service-account activity. If 82% of attacks involve no malware, then identity behavior is the new endpoint. The log of who did what, when and from where is the primary detection surface. Secure the cloud control plane. Centralize logs from cloud, SaaS, identity, endpoints and network infrastructure so movements between them remain visible. The attack moves through the seams. The defense has to watch the seams.
Inventory edge devices aggressively. Patch VPNs, firewalls and gateways quickly. Eliminate abandoned appliances. Verify that logs are actually reaching the SOC. The security device you forgot about is the one the attacker will remember.
Put AI systems into the security model. Inventory agents, models, connectors, API keys, data sources, notebooks and AI-development platforms. Apply least privilege and record what the systems do. If you do not know what AI infrastructure you have, you cannot protect it. And if you cannot protect it, an attacker will use it. Assume credentials will be compromised. Use phishing-resistant MFA, conditional access, device trust and rapid session revocation. Ordinary push-based MFA is increasingly inadequate. When the fastest breakout is 27 seconds, the authentication method that takes a human 10 seconds to approve is not a control. It is a formality.
Automate containment. When breakout can occur in minutes — or seconds — an alert waiting in a queue is not a defense. The response has to be faster than the attack. That means automated containment, automated isolation, automated session revocation. Humans review after the containment, not before.
My read on it
The report's most important finding is not that AI hackers have arrived. It is that the architecture of enterprise security is becoming obsolete. We built defenses around the idea that an attacker would introduce something visibly malicious into the environment. But the modern attacker increasingly uses the organization's own identities, tools, cloud services, software suppliers and now AI systems. They do not need to look like an intruder. They need to look like business activity. And AI does not have to invent a magical new exploit to transform the threat landscape. It merely has to turn a 90-minute attack into a 29-minute attack, allow one operator to conduct ten campaigns instead of one, and make every phishing message, fake employee and malicious script slightly more convincing. That is enough.
Security is shifting from detecting bad software to detecting bad intent moving through legitimate systems — and the time available to recognize that intent is collapsing.
Twenty-nine minutes. Twenty-seven seconds. Four minutes.
The attack is faster. The perimeter is identity. The tools are already inside. And the defender's advantage — the time between breach and detection — is disappearing.
This is the same story I keep finding in every domain AI touches. The execution layer changes, the organizational structure built around the old execution layer becomes the most expensive and least effective part of the operation, and the gap between what the system was designed to defend and what actually needs defending grows until something breaks.
In education, the artifact is no longer evidence of learning. In work, the team is no longer the unit of execution. In security, the malware is no longer the signal of compromise.
The pattern is the same. The map was drawn for a world that no longer exists. And the people holding the map are still pointing at the wrong destination.
Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital.