Attackers are beginning to automate more of the vulnerability exploitation process, according to Google Threat Intelligence Group (GTIG) and Mandiant. Their researchers identified exploit_agent, a multi-agent framework that can inspect software changes after security patches, find vulnerabilities, generate exploits and run them without an operator directing each step.
The discovery marks the first time Google says it has observed threat actors using an AI system to find and exploit flaws autonomously. The framework was found on an exposed staging server and connects to multiple large language models. Its apparent focus on “n-day” vulnerabilities could give attackers a window to target known flaws while organizations are still applying fixes.
Evidence of compromise, but attribution remains unclear
Google found stolen victim data on the server, as well as scripts intended to support activity after an initial breach. The operator showed particular interest in French websites connected to media, entertainment and education. Researchers did not establish whether exploit_agent’s autonomous activity caused the data theft.
A separate campaign illustrates how quickly AI-assisted exploit work can be followed by attacks. Research by GreyNoise, cited in Google’s report, indicates that a threat actor likely used AI tools to develop and test exploits for vulnerabilities in PaperCut print management servers. The campaign targeted hundreds of organizations four days after the flaws were disclosed.
Google also observed two other attacker-used frameworks. CAIRN appeared in intrusion activity targeting organizations that included the US aerospace sector. A customized platform based on PentAGI was used for reconnaissance and vulnerability scanning against organizations across Asia.
Faster attacks demand faster response
The techniques themselves are familiar. Jonathan Ong, a senior analyst at Omdia, described the frameworks as penetration-testing and red-team tools that use established methods, with large language models speeding up the work. Keith Prabhu, CEO of Confidis, said there’s limited evidence that AI makes any one attack more successful. Its broader effect may be to help attackers launch more attempts in less time.
That increase in pace is the operational concern. Ong said frameworks often generate enough detectable activity for well-configured and monitored endpoint detection and response (EDR), extended detection and response (XDR) or managed detection and response (MDR) systems to catch much of it. Google cautioned that detectable activity doesn’t prevent these tools from achieving their objectives.
Ong recommends maintaining established safeguards such as multifactor authentication, strong passwords and controls against repeated login attempts, alongside stronger cloud and identity security. His central point is response speed: organizations need to contain incidents faster as automated attacks move through familiar steps more quickly.
Prabhu pointed to fragmented security visibility and manual response decisions as sources of delay. He recommends continuous asset discovery, risk-based vulnerability remediation and detection that connects suspicious activity across the enterprise. Amit Kumar Jena, development manager for AI at Kanerika, says internet-facing systems deserve a fast patching path, with virtual patching used while permanent fixes are being tested.
Teams can also predefine which containment actions are safe to automate. Prabhu recommends automating reversible actions while keeping human approval for decisions with serious operational consequences. Jena cited endpoint isolation and session revocation as examples of actions organizations can authorize in advance. The aim, Prabhu said, is to remove avoidable delays while testing that response can keep pace with an attack.
More access to automation, uneven exposure
Open-source agent frameworks could help less-experienced criminals coordinate existing tools and adapt activity based on results, Prabhu said. They don’t remove the need for technical skill or guarantee a successful breach, but they could lower the effort needed to attempt more attacks.
Smaller organizations and service providers with limited security resources may be especially exposed. Risk also cuts across industries: poorly maintained legacy systems and weak identity controls can leave organizations vulnerable regardless of sector. For security teams, the practical priority is to shorten the time between a public flaw, a suspicious signal and a containment decision.