Google identifies autonomous AI systems conducting vulnerability attacks

Google has identified cyber attackers using artificial intelligence systems to discover software vulnerabilities and launch exploits without human intervention, marking a new stage in the automation of cyber intrusions.

The Google Threat Intelligence Group (GTIG) disclosed the finding in its quarterly AI Threat Tracker published on 16 November, describing operational systems that monitor software updates, identify weaknesses exposed by security patches and automatically develop attack code.

Investigators discovered an exposed staging server containing a multi-agent framework called exploit_agent, configured to work with both advanced commercial and locally operated language models. The system compared software code before and after security fixes, assessed potential vulnerabilities, generated exploits and executed them autonomously.

The server also contained data taken from targeted victims and scripts supporting further malicious activities, including internal network reconnaissance, password attacks and the installation of webshells, which provide remote access to compromised machines.

Google said information recovered from the system indicated particular interest in French websites connected with media, entertainment and education. Its findings established the presence of an autonomous exploitation pipeline, although the company did not specify how many successful intrusions could be attributed directly to the framework.

The development sharpens concern over the shrinking interval between disclosure of a software flaw and attempts to exploit organisations that have not installed available patches. Rather than requiring specialists to analyse each update manually, automated systems can continuously compare code versions and test possible attacks.

The research distinguishes exploitation of known, patched vulnerabilities from the discovery of previously unknown flaws. Google observed attackers concentrating on weaknesses that could be inferred from released fixes, where accessible code and technical documentation provide a narrower research target.

GTIG and Google’s Mandiant cybersecurity division also documented the use of publicly available multi-agent penetration-testing frameworks during active intrusions. Tools identified included ARTEX, CAIRN, PentAGI and CyberStrikeAI, although the extent of their observed deployment varied.

CAIRN was used in intrusion activity against targets including the United States aerospace sector, Google said. Separately, investigators tracked an attacker using a customised system based on PentAGI to scan organisations across Asia between late August and early October.

Such frameworks coordinate specialised AI agents to perform reconnaissance, identify security weaknesses, produce individual attack payloads and modify tactics according to feedback. They can accelerate familiar attacks involving stolen credentials and unpatched servers without necessarily introducing fundamentally new methods.

The company nevertheless cautioned that current automated attacks often sacrifice concealment for speed. Repeated scans, identifiable software signatures and machine-generated code can leave evidence that helps defenders detect malicious operations.

Google also described how the same technology has accelerated exploitation following vulnerability disclosures. Two flaws affecting PaperCut print management servers were exploited as zero-days in August. Following their disclosure, security researchers produced working exploits using language models within a day.

Cybersecurity company GreyNoise subsequently identified a campaign targeting hundreds of organisations four days after the disclosure. Its analysis suggested the attacker probably used AI tools to prepare exploits and target lists, although that assessment did not establish that every stage was autonomous.

The broader growth in AI-generated exploit material has complicated verification. Google cited vulnerabilities affecting PHP for which large numbers of purported proofs of concept appeared after patches were issued in October, many of them non-functional.

John Hultquist, chief analyst at GTIG, said attackers were generally exploiting established weaknesses rather than inventing entirely new intrusion techniques. The distinguishing change was the speed and scale at which systems could pursue those opportunities.

Google said its security operations division had introduced behavioural and traffic-volume detection rules directed at ARTEX activity. Mandiant has also published defensive guidance addressing the compressed response times associated with automated attacks.

Beyond direct exploitation, the tracker documented attackers stealing credentials and cloud computing resources to operate AI services without authorisation. Mandiant investigated incidents in which fraudulent model usage generated charges exceeding $100,000 in a single day, with cumulative costs reaching $1 million in the most severe cases.

Investigators also observed illicit services reselling access to stolen AI accounts and computing capacity. Some operators distributed requests across compromised environments, making unauthorised usage harder to identify through individual accounts.

Google said attackers deploying autonomous frameworks could still be undermined by unreliable model outputs, including fabricated technical details. The assessment noted that such errors and conspicuous activity could limit their usefulness for espionage operators seeking to maintain undetected access over extended periods.



Notice an issue?

Arabian Post strives to deliver the most accurate and reliable information to its readers. If you believe you have identified an error or inconsistency in this article, please don't hesitate to contact our editorial team at editor[at]thearabianpost[dot]com. We are committed to promptly addressing any concerns and ensuring the highest level of journalistic integrity.


Loading next story…