Google GTIG Finds AI Agent That Builds and Runs Exploits Without Humans
| |

Google GTIG Finds AI Agent That Builds and Runs Exploits Without Humans

AI hacking has entered an automated stage that no longer needs a human hand. In a report released on the 16th (local time), the Google Threat Intelligence Group (GTIG) said it had, for the first time, identified ‘exploit_agent,’ a multi-agent framework that compares code before and after a security patch to build attack code and then executes it without human involvement. According to the report, the number of exploited zero-day vulnerabilities rose from a monthly average of 8 in 2025 to 11 in 2026.

The report concluded that attackers are moving “beyond manual prompting to deploying autonomous, multi-agent attack frameworks.”

exploit_agent

GTIG discovered the framework on an externally exposed staging server (a server used for testing before actual attacks). exploit_agent uses both frontier large language models (LLMs) and locally run LLMs. When a security patch is released for a popular web application, it continuously runs ‘differential analysis’ comparing the code before and after the patch to identify the fixed vulnerability, writes an exploit, and then runs it automatically “without human intervention.” An exploit is attack code that actually takes advantage of a vulnerability.

The server also held scripts for internal reconnaissance, brute-force attacks, web shell installation and removal of web shells planted by other attackers. A web shell is a malicious file secretly planted on a web server to issue commands remotely. Telemetry from the victims’ side suggests the operator targeted French media and entertainment sites and education-sector sites. A configuration file contained Russian-language comments, which GTIG translated as “enter the key directly—right here, not in the chat.”

What makes this case significant is that it targets vulnerabilities that have already been disclosed, so-called ‘n-days.’ If attack code can be generated automatically by reading patches, the gap before administrators apply a patch itself becomes an attack window.

Attacks within days of patch release

The report cited cases in which the time from disclosure to exploitation has shrunk. In late August, two zero-days (CVE-2026-81578 and CVE-2026-82078) in the PaperCut print management server were exploited. Chained together, the two vulnerabilities allow remote code execution without authentication. Within a day of disclosure, researchers and the security firm Huntress built working exploits. The security firm GreyNoise reported that four days after disclosure, attacks apparently assisted by LLMs targeted hundreds of organizations, and in one case the attacker escalated to domain administrator privileges within minutes of the initial intrusion.

Two PHP vulnerabilities (CVE-2026-109618 and CVE-2026-109619) were discovered through LLM analysis. A proof-of-concept (PoC) video and the patch code were made public before the official patch on October 21, allowing attackers to compare versions. Hundreds of non-working PoCs then flooded in; GTIG judged that many were generated by LLMs and noted that they increased the verification burden on defenders.

According to the report, vulnerability disclosures and exploits nearly doubled between January and August this year. Exploited zero-days rose from a monthly average of 8 in 2025 to 11 in 2026, spiking to 22 in August.

Multi-agent attack tools and ‘LLMjacking’

CAIRN, an open-source penetration testing tool in which agents collaborate through an ‘observe-orient-decide-act (OODA)’ loop, was observed in attacks targeting the US aerospace sector. A platform modified from the open-source PentAGI systematically conducted reconnaissance on Asian organizations from late August to early October. CyberStrikeAI connects LLMs to more than 100 security tools, and HexStrike AI connects more than 150 attack tools through more than 12 sub-agents. However, GTIG assessed that these tools “currently favor speed over stealth,” leaving many traces.

‘LLMjacking’ is a technique of stealing and using others’ LLM credentials and cloud resources, or reselling them through ‘LLM mixers.’ In incidents handled by Mandiant, a Google-owned security firm, fraudulent charges exceeded $100,000 per day, and in severe cases cumulative damages reached $1 million. Mixer prices ranged from $0.10 to $25 per million tokens.

State-backed groups and AI coding tools

One contractor combined Google’s Gemini API with other vendors’ tools to build a platform managing more than 100,000 accounts on major platforms, and planned to sell it to government agencies and overseas customers. Google said it had blocked the related assets and updated its models.

The state-backed groups named were Iran’s CALANQUE ION, the Chinese contractor LEAKY CASTLE and the North Korea-linked MIDNIGHT NEPTUNE. MIDNIGHT NEPTUNE targeted the cryptocurrency industry using fake recruiter profile images bearing Google’s SynthID watermark. It then modified the ‘.claude/settings.json’ file in an internal repository so that running Claude Code, Anthropic’s coding tool, would download the SOMBERMEME backdoor. The case shows that configuration files for AI coding tools can become a channel for supply-chain attacks.

The core of the report is that attack automation has changed the question of time. If agents shorten the time attackers need to prepare after a patch is released, defenders’ time to apply the patch shrinks accordingly. However, the report reflects Google’s own observations, and it did not disclose how many patches exploit_agent analyzed or what its attack success rate was. As defensive measures, Google introduced detection rules for its security operations platform, Mandiant’s hardening guidance, and its vulnerability-finding and fixing AI tools ‘Big Sleep’ and ‘CodeMender.’

Source: GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold (Google Cloud)

Related articles

사이언스모니터 뉴스레터
AI·양자·R&D 핵심 해설을 주 2회(화·금) 이메일로 받아 보세요.

Similar Posts