Enterprise patching is becoming a faster operational race as attackers apply agentic AI to vulnerability research and exploitation. Google (NASDAQ: GOOGL) says it has observed threat actors use multi-agent frameworks and LLM-driven workflows to accelerate multi-stage intrusions, including a system that could move from patch analysis to exploit execution without a human in the loop.

The company’s Google Threat Intelligence Group published the findings on Nov. 16 in its AI Threat Tracker. The research does not establish that every organization faces autonomous attacks today, but it documents a practical change in the threat model: attackers can use automation to compress the interval between a public patch and a usable exploit, reducing the time available for enterprises to assess, test and deploy fixes.

Google said one framework, which it calls exploit_agent, was configured to perform continuous differential analysis of popular web-application software. According to the company, the system compared code before and after patch announcements, identified potential flaws, drafted exploits and executed them automatically. Google found the framework on an exposed staging server and said associated telemetry indicated interest in organizations in France’s media, entertainment and education sectors.

The report also describes open-source, multi-agent offensive frameworks including CAIRN, CyberStrikeAI, HexStrike AI and PentAGI. Google said it had observed CAIRN and PentAGI in intrusion activity and had seen attackers experiment with other tools. These systems can combine reconnaissance, vulnerability scanning, exploit development and execution feedback, allowing operators to automate discrete phases of an attack chain rather than relying only on manual prompting.

For security teams, the immediate business implication is not that patch management can be replaced by an AI control. It is that the operating assumptions behind patch prioritization need to account for faster exploit development. Internet-facing applications, known vulnerabilities with available fixes and systems where a patch exposes useful code changes deserve particular attention. Teams also need to connect vulnerability intelligence with detection engineering, because the company said autonomous tooling can leave recognizable traffic patterns and code-level artifacts even as it increases attack velocity.

That requires decisions outside the vulnerability-management console. Application owners need current inventories of externally reachable services and clear escalation paths when an exploited flaw affects a business-critical system. Security operations teams need telemetry that can identify unexpected reconnaissance, repeated exploit attempts and anomalous use of privileged credentials. Change-management processes must still protect production stability, but the research reinforces the cost of queues that leave a known, internet-exposed weakness unresolved while attackers can rapidly test newly available patch information.

Google also reported that criminal groups are monetizing stolen LLM credentials and cloud resources through proxy services, while espionage actors use LLMs for target research, data analysis and potential supply-chain attacks. The company said attackers continue to exploit familiar weaknesses such as compromised credentials and unpatched systems; the change is the speed and scale available when agent frameworks chain tasks together.

That distinction matters for CIOs and security leaders balancing tool investments against operational discipline. The report supports more urgency around asset visibility, rapid remediation, identity controls and monitoring of exposed services, but it does not prove that any single security product will stop agentic attacks. Legacy programs often treat scanning, patching and detection as separate workstreams. Google’s findings suggest attackers increasingly do not: their workflows can connect those stages continuously, making the enterprise response time between them more consequential.