CrowdStrike's "2026 Threat Hunting Report," published on August 3, 2026, along with a case reported by Palo Alto Networks' threat intelligence team Unit 42 on July 30, demonstrate a fundamental shift in the role of artificial intelligence (AI) in cybersecurity. AI has moved beyond simply being a tool that assists attackers. It has become a "weapon" that operates autonomously, and at the same time, it has become the "target" most sought after by attackers.
Over the past year, malicious activity leveraging AI has increased by 89%. This figure confirms that attackers are scaling up and accelerating their attacks by incorporating AI. In proportion to the speed at which enterprises are integrating AI into their businesses, attackers have begun targeting AI infrastructure and compromising software supply chains. Defenders now face a wave of automated threats advancing at unprecedented speed and scale.
The Era of AI Autonomously Hunting Targets
The activities of the Chinese-speaking threat actor "knaithe (also known as KnYuan)" identified by Unit 42 highlight the reality of autonomous attacks. They attempted to automate attacks by embedding the AI model "DeepSeek" as an inference engine into the open-source framework "Hermes Agent." Upon receiving instructions via Telegram, this AI agent would enumerate targets, investigate vulnerabilities, download exploits, and execute attacks—all without human intervention.
The behavior of this autonomous agent was remarkably organized. For instance, in attacks targeting vulnerabilities in Langflow and the workflow automation tool n8n, the agent used the search engine FOFA to discover assets exposed on the internet and autonomously searched GitHub for the latest proof-of-concept (PoC) exploit code. DeepSeek autonomously evaluated factors such as the number of deployments and CVSS scores (severity) of the target, demonstrating the flexibility to immediately pivot to a different target if it judged the value to be low.
In practice, this actor targeted more than 460 systems. At this stage, cases of complete compromise appear to have been limited, as auto-login was disabled in some Langflow environments and authentication was configured on n8n forms. However, the fact that AI autonomously handled everything from target selection to exploit attempts—managing resources and completing tasks in minutes that would normally take hundreds of hours of manual work—carries significant weight. This actor succeeded in data exfiltration and command execution through manual attacks targeting Citrix NetScaler and Marimo Notebook, and if autonomous attacks learn from these techniques and improve their precision, the scale of damage could expand dramatically. The technical barriers are steadily falling.
Collapsing Patch Windows and a Flood of Alerts
The automation and autonomy of attacks have completely erased the time margin that once remained for defenders. According to CrowdStrike's report, exploitation begins within 48 hours for a full 88% of vulnerabilities for which a PoC has been published. Furthermore, Chinese threat actors such as Vault Panda and Genesis Panda have launched deliberate attacks within 24 hours of a vulnerability's disclosure.
The "patch within 30 days" benchmark that many organizations have long targeted has now become a relic of the past. Defenders today are forced to respond within an extremely short patch cycle of just 24 to 48 hours. The day after a vulnerability is disclosed, attacks using AI-generated scripts and commands are already spreading across the internet.
Further compounding the burden on defenders is the sheer volume of noise generated by AI. According to observations by CrowdStrike's threat hunting team, they process an average of 14 million detection leads per day, and detections attributable to AI agents now reach 2.5 times the number of detections caused by humans. Because attackers are deploying scans and initial attacks at massive scale and indiscriminately using AI, defenders must find genuine threats buried within a flood of alerts. The risk of being overwhelmed by the sheer volume of AI-generated signals is growing.
AI Infrastructure and the Supply Chain as Targets
AI is not only a powerful weapon for attacks but has also transformed into a high-value target for attackers. Particularly notable is the emergence of a technique called "LLMjacking," which involves hijacking access itself to large language models (LLMs). In one attack, a threat actor seized administrative privileges over a cloud provider's foundation model service and sent nearly 200,000 API requests within two minutes. There have also been observed cases of cost harvesting, in which token thieves abuse infrastructure to deliberately inflate a company's AI usage costs.
Development environments and the open-source supply chain have also become new battlegrounds within the AI ecosystem. The North Korea-linked threat group Famous Chollima leveraged AI-generated websites, GitHub accounts, and email infrastructure to establish fictitious companies, facilitating infiltration into target organizations. Their affiliated group, Stardust Chollima, stole credentials from npm package maintainers and injected malicious dependencies into 131 packages of the Mastra AI framework. In this campaign, they used LinkedIn video calls to contact employees, cleverly tricking them into clicking malicious links.
Additionally, the financially motivated crime group Altered Spider compromised more than 300 software dependencies in a single day within one campaign. Using stolen maintainer credentials, they executed credential-stealing malware within automated build pipelines. If developers incorporate these tainted packages while adopting AI frameworks and tools, deep intrusion into cloud environments can occur within minutes.
A New Defensive Battle Where Response Speed Is the Deciding Factor
While enterprises rush to adopt the latest AI to maintain competitiveness, that very AI infrastructure is creating an attack surface with insufficient defenses. Beyond simply leveraging AI as a defensive weapon, organizations must also address how to protect AI infrastructure itself and the supply chains that accompany it.
As attackers continue refining autonomous agents and exploiting vulnerabilities one after another at machine speed, defenders are now required to embed security measures at a speed equal to or exceeding the pace of AI adoption. Manual alert reviews and patch cycles measured in weeks are no longer viable countermeasures. Strengthening monitoring of AI infrastructure and building automated defense and response mechanisms will constitute the minimum baseline of defense in the era ahead.