Google said it has identified a hacker-operated AI framework designed to discover software vulnerabilities and execute exploits without human intervention, marking the first time its researchers have observed such a system being used by threat actors. The findings were published in a report by Google Threat Intelligence Group (GTIG) and Mandiant on Monday.

Source: CSO Online

The framework, called exploit_agent, is a multi-agent system that connects to multiple large language models (LLMs) — frontier and locally hosted models, according to the coverage — and continuously compares popular web application code before and after each security patch. From the difference it derives the weakness the patch closed, writes an exploit for it, and executes that exploit against targets, repeating the cycle with no human reviewing the individual steps. GTIG rated the activity at Level 5 autonomous capability in its maturity timeline, coverage of the report said; two months ago, the group's previous report still said it had not seen fully autonomous attack pipelines deployed in the wild.

"A significant threshold" — GTIG

GTIG describes the finding as the first instance of threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop — a step beyond AI helping human hackers with individual tasks.

Found on an exposed server, aimed mainly at French websites

Google's researchers discovered exploit_agent on an exposed staging server. The server also held data exfiltrated from targeted victims, alongside internal reconnaissance tools, web shells and brute-force scripts — signs, investigators say, that the tooling had been used against live targets rather than sitting as an experiment. Telemetry suggested the operator focused on French websites in the media, entertainment and education sectors. The report did not establish whether the framework's own autonomous operations were responsible for the theft of the data found on the server.

Source: CSO Online

Level 5GTIG's autonomous-capability rating for the activity
Nov 16GTIG and Mandiant report published
~3 monthsGTIG expectation for attacks needing little to no human interaction (via BleepingComputer)
~6 monthsGTIG expectation for sophisticated fully autonomous attacks (via BleepingComputer)

exploit_agent was not the only agentic framework in the report. GTIG said it has observed CAIRN deployed in intrusion activity involving US aerospace targets, and a customised platform based on PentAGI conducting reconnaissance and vulnerability scanning against organisations across Asia between late August and early October, with further experimentation involving CyberStrikeAI and HexStrike AI. In a separate campaign cited from GreyNoise research, a threat actor likely used AI tools to develop and test exploits before launching a campaign against hundreds of organisations four days after vulnerabilities in PaperCut print management servers were disclosed — the shrinking patch-to-exploit clock that defenders fear.

Noisy, not invisible — and no operator named

GTIG also adds a caveat: these frameworks currently trade stealth for velocity. They generate noisy traffic and leave code-level traces, so good logging and monitoring can still catch them. Google has not named the operator or any state behind exploit_agent; a configuration file carried Russian-language comments, and coverage described a suspected Russian-speaking operator. In one worked example, the report shows the framework inspecting access-control code in the Spectra (Ultimate Addons for Gutenberg) WordPress plugin, version 2.19.25 — it does not say a flaw there was confirmed or exploited.

What is solid

  • GTIG found the exploit_agent framework on an exposed staging server and described it in its Nov 16 report with Mandiant
  • The framework compares code before and after patches, generates exploits and runs them without a human directing each step, according to GTIG
  • Data exfiltrated from victims was present on the same server, and targets pointed mainly to French media, entertainment and education websites

What is only claimed

  • Whether exploit_agent's own exploits stole that data — the report did not establish this
  • Who operates it — Google named no group or state; Russian-language comments only suggest a Russian-speaking operator
  • Whether the Spectra plugin code it inspected contained an exploitable flaw — not confirmed in the report

Analysts quoted in the coverage framed the shift as speed, not magic. "This is more of an acceleration of existing attack techniques," said Jonathan Ong, a senior analyst at Omdia, noting the frameworks build on established pentesting tools sped up by LLMs. Keith Prabhu, CEO of Confidis, said the greater risk is more attempts in less time, raising the number of successful compromises even if each attempt is no more effective. The practical defence advice is unglamorous and urgent: patch internet-facing systems fast, harden cloud and identity controls, use multifactor authentication, and automate the containment steps that are well understood and reversible.

Source: CSO Online

What happens next will be measured in patch windows. If GTIG's forecast is right — simple attacks needing little to no human interaction within about three months, sophisticated fully autonomous attacks within about six — then the days between a vendor publishing a patch and an attacker weaponising it, already down to single digits in cases like the PaperCut campaign, become the decisive battleground for every organisation running public-facing software.