Google’s Threat Intelligence Group has issued a warning that cyber‑criminals are now employing artificial‑intelligence agents to automatically locate and exploit vulnerabilities revealed by software patches. The briefing, titled “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold,” describes a multi‑agent system called exploit_agent that can conduct the entire attack workflow without human direction.
How the AI framework operates
The reported framework functions as a continuous scanner of publicly released software updates for widely used applications. By comparing the source code before and after a patch, the system isolates the specific weakness that the vendor has remedied. Once the fixed flaw is identified, the AI proceeds to craft a targeted exploit aimed at organizations that have not yet applied the new version.
Google says the framework was discovered on an exposed staging server that was linked to several frontier‑grade and locally hosted large language models (LLMs). These models supplied the reasoning capability needed to interpret code differences, generate exploit code and orchestrate follow‑up actions such as reconnaissance, brute‑force attempts, and web‑shell deployment.
A configuration file on the server contained a Russian‑language note directing the operator to insert LLM API keys directly, underscoring the automated nature of the setup. The server also stored data harvested from compromised victims, along with scripts for internal probing, installing and removing malicious web shells, and other post‑exploitation tasks.
Evidence of exploitation and impact
The GTIG team observed that the AI system focused on “n‑day” vulnerabilities—flaws that have already been patched but remain exploitable on unpatched systems—rather than hunting for previously unknown “zero‑day” bugs. A patch release supplies a finite set of artifacts, including pre‑ and post‑patch source, release notes, security analyses and proof‑of‑concept code, all of which the AI can ingest and repurpose.
Two specific weaknesses in PaperCut print‑management servers were highlighted as case studies. Both were exploited as zero‑day attacks in late August 2026. Within a day of public disclosure, security researchers and the firm Huntress were able to generate functional exploits using LLMs. Four days later, GreyNoise identified a campaign that appeared to have leveraged the same AI‑driven process to build and test exploits before rolling them out against hundreds of targets.
One documented intrusion demonstrated the speed of the approach: after gaining an initial foothold, the attacker escalated to domain‑administrator privileges in a matter of minutes. Victim telemetry indicated a particular interest in French‑based websites operating in the media, entertainment and education sectors.
Industry response and future outlook
John Hultquist, chief analyst for Google’s GTIG, emphasized that the underlying techniques are not novel. “Attackers are still chasing known, unpatched issues, but they are doing so at a dramatically faster pace and at a larger scale,” he said. He warned that traditional, manually‑driven defenses will struggle to keep up with machine‑speed threats.
Hultquist urged organizations to adopt AI‑enabled defensive measures, describing the need for “agentic defence” to avoid falling behind adversaries that have already automated their offensive playbooks. Google confirmed that it is applying its own AI models to discover and remediate vulnerabilities before malicious actors can weaponise them.
The report arrives amid a broader wave of AI‑related cyber incidents. Earlier this year, a breach of the cryptocurrency exchange Binance was attributed to attackers exploiting an internal AI tool, a compromise that cost an estimated $6.3 billion. Microsoft recently disclosed that a China‑linked group used an AI framework named ARTEX to launch automated attacks against hundreds of organizations worldwide. India’s computer emergency response team, CERT‑In, has also issued warnings about the growing use of AI tools by threat actors.
Collectively, these developments suggest that the cybersecurity landscape is shifting from isolated, human‑crafted exploits to fully automated, AI‑driven attack chains. Companies that continue to rely solely on manual patch management and signature‑based detection may find themselves outpaced by adversaries capable of generating, testing and deploying exploits in hours rather than weeks.
Google’s findings underscore the urgency of integrating artificial intelligence into both threat‑intelligence gathering and active defense. By automating the identification of vulnerable code, correlating patch information and generating exploit payloads, AI agents can dramatically reduce the window of opportunity for attackers. The message from the tech giant is clear: to stay resilient, defenders must match the speed of the machine.






Be First to Comment