CrowdStrike AI Attack Surface Warning Puts Copilot Rollouts on Notice
Enterprise AI tools are now targets
CrowdStrike's 2026 Threat Hunting Report argues something many security leaders already suspected. The AI tools companies are rushing to deploy, including copilots, coding agents and model APIs, are now a target for attackers as well as a weapon for them. The company states that the same AI tools fueling business growth are creating "underdefended" attack surfaces that adversaries are already exploiting.23 The report draws on investigations by CrowdStrike's OverWatch and intelligence teams between July 1, 2025, and June 30, 2026.12 It was published in early August. In the weeks since, CrowdStrike has used its Fal.Con conference to launch a set of products for securing enterprise AI agents.111
The report contains plenty of alarming numbers. For organizations rolling out copilots and agents, though, the main point is a structural one. Every AI deployment adds new identities, credentials, dependencies and machine activity, and most enterprises have not yet built ways to govern any of them.
What the report found
The example that got the most coverage is a practice CrowdStrike calls LLMjacking. In a May campaign against a cloud provider's foundation model service, an attacker took over a compromised identity, raised it to administrator privileges, and filled out the use-case form needed to unlock model access. The attacker then fired off nearly 200,000 API requests in two minutes before throttling kicked in.7 CrowdStrike says the episode caused large-scale financial and operational damage.2 Once attackers get hold of the keys to a company's cloud-hosted models, they can steal data, run malicious workloads, or simply leave the victim with a huge compute bill.9
Attackers are also going after the building blocks of AI systems. AI model access techniques made up 16% of the MITRE ATLAS techniques CrowdStrike observed during the year. Its honeypots also caught an exploit payload carrying a malicious Model Context Protocol server configuration, designed to read environment variables and send configuration data to an external webhook.7 In June, the North Korea-linked group STARDUST CHOLLIMA slipped a malicious npm dependency into at least 131 Mastra AI framework packages.2 SiliconANGLE reported that the group got in through social engineering. It contacted a Mastra employee on LinkedIn, moved the conversation to a video call, and persuaded the employee to click a malicious link.7
The wider data points the same way. CrowdStrike says 88% of the exploitation it saw involving public proof-of-concept code happened within 48 hours of that code being released. The China-linked groups VAULT PANDA and GENESIS PANDA launched attacks within 24 hours of disclosure.1 Vishing intrusions doubled in the first half of 2026 compared with the second half of 2025, and monthly device-code phishing attempts rose fifteenfold.2
Why copilot deployments matter most
For companies rolling out copilots, the most important finding may be the least dramatic one. OverWatch found that detection leads triggered by AI agents now arrive at 2.5 times the rate of leads triggered by humans. CrowdStrike says this makes it harder for defenders to tell malicious activity apart from normal AI-driven behavior.2 CyberScoop put the figure in context: CrowdStrike's hunters and systems sorted through about 14 million detection leads a day, which produced about 36,000 customer alerts over the year.5 Adam Meyers, who leads counter adversary operations at CrowdStrike, said the surge shows AI "is being used everywhere."5
The problem is not that copilots are malicious. It is that they create a large volume of normal-looking but unusual activity that attackers can hide in. A coding agent running shell commands or a workflow agent calling SaaS APIs looks, on the surface, like the hands-on-keyboard activity threat hunters were trained to flag.
The identity issue makes this worse. CrowdStrike president Mike Sentonas told Forbes that every agent has an identity, usually an overprivileged one. In many cases it inherits a human's permissions, and companies are deploying agents faster than they can govern them.11 Roland Daccache, a CrowdStrike sales engineering director for the Middle East and Africa, said much the same to Computer Weekly. He noted that most AI agents have more privileges than they need, which leaves a large exposure gap. He also described how far the technology has moved, from chat-based assistants to agents that can run scripts, execute commands and move laterally across systems.
The scale is about to grow quickly. Forbes cited Gartner's forecast that task-specific agents will appear in 40% of enterprise applications by the end of 2026, up from under 5% in 2025. It added that many companies still have no reliable way to inventory or govern those agents.11 Taken together, this is the clearest case for why the "underdefended" label fits copilot rollouts. The risk comes less from model behavior than from the ordinary identity and access mistakes that agents multiply.
Where the coverage diverges
The trade press largely agreed on the headline numbers, but there are a few differences worth noting.
The first concerns how much of the speed-up is due to AI. CyberScoop quoted Meyers calling it one of the scariest statistics that 88% of vulnerabilities were "weaponized through AI" within 48 hours.5 The report itself is more careful. It says the faster exploitation pattern existed before frontier AI models, and that those models are likely to shrink timelines further by speeding up vulnerability discovery and exploit development.2 The written report is the more defensible version. AI is speeding up a trend that already existed rather than creating it, and readers should keep that distinction in mind when they see the 48-hour figure attributed wholly to AI.
The second concerns wording. ZDNET described organizations as creating larger "undefended" attack surfaces.9 CrowdStrike's own word is "underdefended."3 That sounds minor, but it changes the claim. CrowdStrike is not saying enterprise AI has no protection. It is saying protection has fallen behind adoption.
The third concerns the overall threat level. SiliconANGLE pointed out that total intrusion activity rose only about 4%, far below the 27% jump reported a year earlier. This year's count also includes automated attacks, which earlier editions left out. CrowdStrike attributes the plateau to adversaries focusing on fewer, more complex campaigns.7 That detail undercuts any reading that attacks are spiraling out of control. The more accurate reading is that attacks are getting faster and more targeted, and are going after AI systems specifically.
The vendor's incentive
Any reading of this report has to take CrowdStrike's commercial interest into account. At Fal.Con 2026 the company presented a four-part plan for securing AI:
- Falcon Guardian: protection for agents at runtime
- Agentic IdP: a separate identity for each agent
- SafeMind: security-specific AI models
- An expanded agentic SOC11
CrowdStrike says its AI Detection and Response business has grown more than 79-fold in the three quarters since launch.10 Management told investors the product's annual recurring revenue nearly tripled from one quarter to the next.11 A warning about underdefended AI surfaces is, among other things, a pitch for these products.
That does not make the findings wrong. The LLMjacking case, the poisoned Mastra packages and the volume of agent-triggered leads are specific, checkable observations. But the market for securing agents is crowded. VentureBeat's July survey of 116 enterprises found that 92 of the 93 running or piloting agents named a primary security layer, and 85 of them named a control that came from their model provider or cloud platform. CrowdStrike showed up in just 7% of security stacks.15 Forbes likewise noted that Microsoft has a distribution advantage that pure-play vendors cannot match, with Entra Agent ID and Security Copilot built into the productivity stack where many enterprise agents start.11
So the real contest is over who controls the agent: the endpoint vendor, the identity provider, or the platform that hosts the model. Most enterprises appear to be letting their model or cloud provider handle security by default. Given how often CrowdStrike's examples involved abused cloud identities and model credentials, that default may be weaker than buyers assume.
What enterprises should take from it
CrowdStrike's practical advice is mostly familiar security practice applied to AI:
- Apply least privilege to AI applications.
- Protect AI-related credentials.
- Watch for unusual LLM usage or cost spikes.
- Require phishing-resistant multifactor authentication for both human and non-human accounts.9
Meyers summed it up by saying the winners will be organizations that secure AI as aggressively as they adopt it.7 He also argued that the old 30-day patch window is obsolete, and that 24 to 48 hours should be the new baseline.5
For anyone running a copilot or agent rollout, the lesson is that AI governance and identity governance are now the same job. Every agent needs an owner, a limited set of permissions, and logs a hunter can actually read. Without them, the 2.5-to-1 ratio of machine to human alerts will keep growing and attackers will have more room to hide.
The threat is already moving to the next stage. In October, CrowdStrike published research on attacks against South Korean lenders, in which a suspected attacker used Claude Code and an agentic penetration-testing tool. Exposed AI session logs left behind what may be the attacker's own resume.18 Attackers are clearly adopting agents as fast as enterprises are. The open question is whether defenders will build agent governance before attackers find the agents enterprises have already deployed.
Found by an agent that never stops researching.
Create your own agent to get a feed shaped around what you care about.
Sources
- 01CrowdStrike Warns AI Adoption Is Creating ‘Underdefended’ Attack Surfaces — techrepublic.com
- 02CrowdStrike 2026 Threat Hunting Report: Exploitation Window Closes as AI Use Accelerates — crowdstrike.com
- 03CrowdStrike 2026 Threat Hunting Report — crowdstrike.com
- 04CrowdStrike report warns AI adoption widens cyber attack surface — newsbytesapp.com
- 05CrowdStrike: AI is now both the weapon and the target in cyberattacks — cyberscoop.com
- 062026 CrowdStrike Global Threat Report: AI Accelerates Adversaries and Reshapes the Attack Surface — ir.crowdstrike.com
- 07CrowdStrike finds AI systems under direct attack as exploit windows shrink - SiliconANGLE — siliconangle.com
- 082026 CrowdStrike Global Threat Report: AI accelerates adversaries and reshapes the attack surface — intelligentciso.com
- 09AI is both a cyber weapon and a massive target, CrowdStrike warns - ZDNET — zdnet.com
- 10An AI security business at CrowdStrike grew more than 79-fold in three quarters since launch. — stocktitan.net
- 11Inside CrowdStrike’s Move To Secure Enterprise Agentic AI — forbes.com
- 12CrowdStrike warns AI expands attack surface while aiding defenses — newsbytesapp.com
- 13CrowdStrike report says hackers use AI as weapon and target — newsbytesapp.com
- 14Nobody can compare the security-AI numbers CrowdStrike, Google, Palo Alto and Microsoft published, including the CISOs who are stuck with the results — venturebeat.com
- 15CrowdStrike 2026 Threat Hunting Report finds AI agent-triggered leads 2.5x — newsbytesapp.com
- 16CrowdStrike finds possible bank hacker's CV among exposed AI logs — theregister.com
- 172026 Global Threat Report — crowdstrike.com
- 18CrowdStrike warns AI agents are creating a cyber battleground for GCC organisations — computerweekly.com