deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

Google's AI Threat Tracker logs exploit_agent, first LLM framework to fully automate exploit development

Google's Threat Intelligence Group says the exploit_agent framework autonomously finds bugs, writes and runs exploits, as zero-day attacks and LLM credential theft surge.

Google's AI Threat Tracker logs exploit_agent, first LLM framework to fully automate exploit development

First fully autonomous exploit pipeline documented

Google's Threat Intelligence Group (GTIG) and Mandiant have documented the first observed case of attackers using large language model frameworks to automate the entire path from vulnerability discovery to a working exploit, with no human in the loop. The finding appears in the AI Threat Tracker edition published on 16 November 2026, which dev.io [sic] coverage summarizes under the report's own framing that agentic threat activity has crossed a significant threshold.

The centerpiece is a framework GTIG calls exploit_agent, recovered from an exposed server. It is a multi-agent system connected to both frontier LLMs and locally hosted models. According to the report, it continuously pulls the code of widely deployed web applications, compares each version before and after a security patch, infers which code changes point to exploitable flaws, drafts exploits, and then executes them on its own. The operator of the compromised server directed this activity at French media, entertainment and education websites.

Open-source agentic attack tools reach real intrusions

Beyond the bespoke framework, the tracker confirms that the open-source agentic pentesting frameworks ARTEX, CAIRN and PentAGI have been used in actual intrusions, and that attackers have been experimenting with two further tools, CyberStrikeAI and HexStrike AI. In response, Google SecOps has shipped detection rules covering ARTEX, and Mandiant has published hardening guidance for environments facing LLM-based offensive tooling.

Patch windows are collapsing

The report ties the tooling shift to a measurable acceleration in exploitation. Vulnerability disclosures and exploits nearly doubled between January and August 2026. Average zero-day exploitation rose from eight per month in 2025 to eleven per month in 2026, peaking at 22 zero-days in August 2026. In late August, two PaperCut NG/MF vulnerabilities, CVE-2026-81578 and CVE-2026-82078, were exploited in the wild as zero-days, with LLM-assisted exploits appearing within a day. In one campaign described in the report, an intrusion that hit hundreds of organizations reached domain administrator privileges within minutes.

Stolen LLM credentials become a revenue stream

The tracker also charts the growth of LLMJacking, in which threat actors steal cloud and LLM credentials and resell access through services the report labels "LLM Mixers". In the cases reviewed, fraudulent charges passed $100,000 in a single day, and cumulative resource abuse reached up to $1 million in the worst instances.

A related tactic targets developers directly. Attackers are tampering with the configuration files of AI coding assistants, with the report citing .claude/settings.json as an example, planting backdoor downloads that trigger when developers launch the tools.

Why it matters

The exploit_agent finding removes, for the first time, the human bottleneck in vulnerability research on the attacker's side, and the window between a patch shipping and a functional exploit is now measured in days or hours. A monthly patch cycle therefore leaves internet-facing applications exposed to flaws that have already been fixed and are already being exploited as n-days. The practical consequences for defenders follow directly from the report: patch exposed applications as soon as fixes are released, treat Mandiant's hardening guidance and the new Google SecOps detections for ARTEX as a checklist, and audit cloud and LLM billing for the fingerprints of LLMJacking, such as sudden usage spikes, unknown service accounts and quota increases, rotating any exposed API keys and session tokens immediately. Teams using AI coding assistants should also review agent configuration files and repositories for tampering before running them, since these files have become an initial access vector aimed straight at developer machines.

  • #security
  • #llm
  • #zero-day
  • #google
  • #threat-intelligence

Related posts