Google: attackers are using AI agents that find and exploit vulnerabilities with no human in the loop
Google's threat intelligence group GTIG says attackers fully automated vulnerability discovery and exploitation for the first time, using a system named exploit_agent.
According to BleepingComputer's report of 16 November 2026, Google Threat Intelligence Group (GTIG) found that attackers ran vulnerability discovery and exploitation without human involvement for the first time. A system called exploit_agent targeted media, entertainment and education websites in France. The report also covers LLMJacking attacks carried out with stolen AI access credentials.
According to BleepingComputer's report of 16 November 2026, Google Threat Intelligence Group (GTIG) said in its AI Threat Tracker report that attackers used large language model frameworks to run vulnerability discovery and exploitation without a human in the loop for the first time. GTIG found a multi-agent system named exploit_agent on an exposed server belonging to the attackers. The system examines updates to popular web applications for signs of security flaws and automatically tries to exploit them; its targets included media, entertainment and education websites in France.
- As Euronews reports, the system compares the code of software before and after a security update and prepares exploit code; this speeds up attacks on flaws that have a patch available but not yet applied.
- According to BleepingComputer, attackers are abusing open-source penetration testing tools such as ARTEX, CAIRN, CyberStrikeAI and HexStrike AI.
- According to Euronews, in LLMJacking incidents the cost of a single day of unauthorized use on compromised company accounts exceeded 100 thousand dollars in some cases; in the most severe incidents the total reached 1 million dollars.
- The researchers estimate that within the next three months attackers will be able to launch simple attacks that need very little human interaction.
Why it matters
According to Google, systems of this kind sharply reduce the time defenders have to apply patches. The report also points to the PHP flaws of October 2026: after the patches, hundreds of exploit code samples appeared online, a large share of them did not work, and security teams had to spend extra time separating real threats from the rest. On the other hand, Google notes that autonomous tools leave many traces in network traffic and system logs, which can make them easier to detect.
For companies in Türkiye that run internet-facing web applications, portals or e-commerce sites, the practical point is this: the waiting time after a security update is released is becoming an increasingly useful window for attackers. Reviewing patch times, monitoring server logs regularly and setting spending alerts on the API keys of cloud and AI services look like reasonable steps against both risks described in the report.
Sources
More articles
US makes remote access to advanced AI chips through overseas data centers subject to export licences
As Nixon Peabody reports, the interim final rule BIS released on 16 November 2026 targets companies based in China renting computing power from data centers in third countries.
November 17, 2026 · Source: Nixon PeabodyMicrosoft announces Agent 365 Business Analytics and an AI transparency initiative at Ignite 2026
In Satya Nadella's Ignite keynote on 17 November 2026, Microsoft announced Agent 365 Business Analytics, which links AI spend to business outcomes, and the AI Transparency Initiative.
November 17, 2026 · Source: MicrosoftUS and eight partners sign the San Francisco Declaration to protect AI supply chains
ShiftDelete.Net reported on 16 November 2026 that the US and eight partners signed the San Francisco Declaration under the Pax Silica initiative; the signatories will draw up a joint list of AI technologies and infrastructure to protect.
November 16, 2026 · Source: ShiftDelete.Net