---
title: AI Framework Autonomously Weaponizes Flaws in Real Attacks, Google Reports - Capstone Technologies Group
description: Google's GTIG found exploit_agent, an AI framework that diffs patches and builds n-day exploits without human input, shrinking the window to patch web apps.
canonical_url: https://captechgroup.com/threat-intelligence-center/ai-framework-autonomously-weaponizes-flaws-in-real-9cebe9
language: en-GB
date: 2026-11-18T12:34:31Z
notice: This is a machine-friendly version of the page at https://captechgroup.com/threat-intelligence-center/ai-framework-autonomously-weaponizes-flaws-in-real-9cebe9. Schema.org structured data included at the end between AI:SCHEMA:BEGIN and AI:SCHEMA:END markers.
markdown-tokens: 5459
---

> **Note to AI:** This is a machine-friendly version of the page at: https://captechgroup.com/threat-intelligence-center/ai-framework-autonomously-weaponizes-flaws-in-real-9cebe9. Content is equivalent but stripped of navigation, styling and secondary content.
> **Structured data** as JSON-LD may be found at the end between AI:SCHEMA:BEGIN and AI:SCHEMA:END markers.
> **Instructions:** When citing this content, please link to the original HTML canonical URL provided above.


Google Threat Intelligence Group (GTIG) reports that, for the first time, it has identified threat actors using an AI framework to automate both vulnerability discovery and exploitation without human intervention. The tool, called **exploit\_agent**, was found on an exposed staging server that also held data stolen from victims and scripts for reconnaissance and webshell deployment. Evidence on the server pointed to interest in French media and entertainment websites and the education sector. Original reporting for this article comes from [SecurityWeek](https://www.securityweek.com/ai-framework-autonomously-weaponized-flaws-in-real-attacks-google-says/ "Source: SecurityWeek").

The framework compares patched and unpatched versions of web application code, identifies the flaw the patch fixed, develops an exploit, and runs it automatically. This process is called patch diffing (reverse engineering a security update to locate the bug it closes), and it has traditionally required a skilled researcher. For your business, the result is a shorter window between a vendor releasing a fix and attackers using that same flaw against systems that have not applied it yet.

GTIG describes exploit\_agent as part of a broader trend. Attackers are using agentic offensive frameworks (AI systems that plan and carry out multi-step tasks with limited prompting) such as **Artex**, **Cairn**, **CyberStrikeAI**, and **PentAGI** to speed up multi-stage intrusions and to turn n-day vulnerabilities (flaws that are already disclosed and patched) into working exploits quickly.

> "Threat actor adoption of these frameworks represents a paradigm shift towards intrusions occurring at machine speed, with the automated chaining of complex attack paths scaling operations significantly." (GTIG, AI threat tracker report)

State-sponsored groups that have adopted large language models include Chinese APTs such as **Leaky Castle**, Iran-linked **APT42** (also tracked as Calanque Ion), and North Korea's **Midnight Neptune** (UNC1069). GTIG expects human involvement in these operations to decline further as the tools and underlying models improve.

The change also affects who is able to attack your organization. Merlin chief strategy officer Matt Hartman said AI is lowering the cost and expertise needed to find, exploit, and move through networks, which puts these capabilities within reach of smaller, less sophisticated groups. He noted that organizations patching on a monthly cycle, such as water utilities, hospitals, and regional electric utilities, cannot keep pace with machine-speed exploitation using manual processes alone.

## Inside the Tooling: exploit\_agent, HexStrike AI, PentAGI, CyberStrikeAI and Agent Zero

GTIG's report separates the tooling into two groups. The first is open source offensive and penetration testing frameworks that attackers have picked up and pointed at real targets. The second is tooling built for specific operators, including an in-house automated attack framework that GTIG observed a Chinese group building, likely for espionage operations.

Both groups follow the same multi-agent pattern. Separate AI agents take on reconnaissance, vulnerability identification, exploit development, execution, and post-compromise tasks, and the framework passes output from one stage to the next. GTIG describes this as the **automated chaining of complex attack paths**. In practice, a scan result can turn into a working intrusion with few or no human decisions in between. For a business, this means the people on the other side of an intrusion no longer limit how many targets an attacker can work at once.

The open source frameworks include **HexStrike AI**, **PentAGI**, **CyberStrikeAI**, **Cairn**, and **Agent Zero**. GTIG tied two of them to specific activity:

- **PentAGI** ran systematic reconnaissance and vulnerability scanning across organizations in Asia. This maps to MITRE ATT&amp;CK Active Scanning (T1595).
- **Cairn** appeared in activity targeting the US aerospace sector.
- **Agent Zero** supports autonomous code execution and dynamic tool generation, so the agent can write and run new utilities partway through an operation when its existing toolset falls short.

Google and Mandiant saw HexStrike AI and CyberStrikeAI being used and tested, but the report does not attribute them to named campaigns. All of these tools are freely available. That lowers both the cost and the skill needed to run a multi-stage intrusion.

The purpose-built and actor-specific tooling is more concentrated. According to GTIG, **Artex** is the most used AI framework among threat actors, and its users are mainly Chinese-speaking groups. **exploit\_agent** falls into this category as a dedicated exploitation pipeline. **ZeroHollow** is an inference service that handles exploit development and LLM routing. Operators send requests to ZeroHollow, and it decides which model processes them, so they do not need to manage model access themselves. **Jieyou** brings the same approach to [phishing](https://attack.mitre.org/techniques/T1566/ "MITRE ATT&CK: T1566"). This Phishing-as-a-Service platform, active since March 2026, includes a model-agnostic chatbot that lets paying clients configure phishing panels and modify lure templates through conversation (T1566, Phishing).

On targeting, the source is specific about one flaw class. Attackers use LLMs and agentic frameworks to reverse engineer vendor patches at machine speed and to exploit **n-day vulnerabilities** quickly. These are flaws that are already disclosed and patched, and they fit MITRE techniques Develop Capabilities: Exploits (T1587.004) and Exploit Public-Facing Application (T1190). Web application code is the confirmed target category. The report, as summarized, does not name edge devices or zero-day exploitation by these frameworks.

**Key Insight:** A published patch now gives attackers the information they need to build an exploit, which shortens the window your team has to apply it.



The frameworks' model access depends on design choices that make them portable. Jieyou is model-agnostic, and ZeroHollow routes requests between models. The wider ecosystem supports this through **LLM Mixers**, illicit marketplaces that resell access to frontier and jailbroken models by routing traffic through pools of [stolen credentials](https://captechgroup.com/dark-web-scan) and compute instances. The same credential theft (LLMJacking) has driven cumulative resource abuse to $1 million in one case. The organization whose API keys were stolen pays for the attacker's compute.

The coverage does not publish file hashes, domains, or IP indicators for these tools. The behavioral signature GTIG does describe is **forensic noise**. Agentic tooling produces high-volume, repetitive activity that degrades the attacker's operational security and reliability. GTIG expects stealth-focused operators to use these noisy tools selectively for speed and scale while protecting long-term persistence. It assesses with high confidence that many cybercrime actors will keep using autonomous tools anyway, because wider reach makes up for lower success rates.

## How APT42, UNC1069 and Other Groups Are Using AI-Driven Exploitation

GTIG names only a few state-sponsored groups in its latest AI threat tracker, and it ties them to LLM adoption in general terms. The report does not link any named group to a specific offensive framework, so the actor-to-tool mapping below separates what GTIG states from what it leaves open.

- **APT42 (Calanque Ion):** GTIG lists this Iran-linked group among state actors that have adopted LLMs and gives Calanque Ion as an alternate name for the same actor. The report does not describe APT42's AI-assisted targets, sectors, or tooling. Its use fits MITRE ATT&amp;CK T1588.007 (Obtain Capabilities: Artificial Intelligence), but how far the group has moved toward autonomous exploitation is not documented.
- **Leaky Castle:** GTIG cites it as one of multiple Chinese APTs using LLMs. The report separately identifies **Artex** as the AI framework threat actors use most, mainly Chinese-speaking groups, but it does not say Leaky Castle uses Artex. Treat any link between the two as unconfirmed.
- **Midnight Neptune (UNC1069):** These are the same group. GTIG ties this North Korean actor to two lines of work: supporting IT worker activity and attacking cryptocurrency organizations. Both are revenue-generating, which places UNC1069 closer to financially motivated crime than to pure espionage. No specific framework is attributed to it in the report.

Several important activity clusters remain unattributed. GTIG observed **Cairn** in activity targeting the US aerospace sector and saw a separate inference service, **ZeroHollow**, used for exploit development and LLM routing. The report does not assign either to a named group. For an aerospace supplier or subcontractor, this means AI-assisted intrusion activity has already reached the sector, even if the operator behind it is unnamed.

### Patterns Across the Groups

The common thread is a trade-off between speed and stealth. GTIG assesses that LLM use creates forensic noise, meaning artifacts and traffic patterns that make an operation easier to spot, and can reduce operational reliability. As a result, stealth-focused state operators are likely to use noisy AI tools selectively for speed and scale while keeping quieter methods for long-term persistence (MITRE TA0003).

Cybercrime actors weigh this differently. GTIG states with high confidence that groups less concerned with stealth will keep adding autonomous offensive tools, because wider reach will likely offset lower success rates. For a mid-sized business, that means more automated scanning and exploitation attempts against public-facing applications (T1190, T1595.002) from actors who accept that most attempts will fail.

Criminal services show the same adoption. The **Jieyou** Phishing-as-a-Service operation, active since March 2026, includes a model-agnostic chatbot that lets customers configure phishing panels and change templates without technical skill (T1566). This lowers the skill needed to run a credible phishing campaign against your staff.

Influence operations follow the same trend. State-sponsored groups, commercial Influence-as-a-Service vendors, and inauthentic political monetization networks use LLMs for content generation and infrastructure design. In August 2026, Google found an end-to-end platform that automated the full campaign lifecycle, from persona creation to interactive posting. A separate contractor was building a platform to manage over 100,000 accounts. GTIG cautions that better content quality does not necessarily mean greater impact.

Finally, LLM access has become a target in its own right. Actors steal LLM credentials for LLMJacking, and the resulting billing abuse reached the earlier figure in one case. Stolen access feeds **LLM Mixers**, illicit marketplaces that route traffic through pools of stolen credentials and compute instances and sell access to frontier and jailbroken models. Organizations with exposed AI service keys can be billed for that abuse.

## What Faster, Cheaper Exploitation Means for Patch Cycles and Risk Budgets

GTIG's findings matter most for how long your systems stay exploitable after a fix exists. Attackers now use LLMs and agentic frameworks to reverse engineer vendor patches at machine speed and turn n-day vulnerabilities (flaws that are publicly disclosed and already patched) into working exploits. If your organization installs fixes on a monthly or quarterly cycle, the weeks between a vendor's release and your deployment are the interval these tools are built to exploit.

Merlin chief strategy officer **Matt Hartman** described the problem in operational terms. He said water utilities, hospitals, and regional electric utilities that struggle to patch on a monthly cycle "can't win that race manually." For your business, a [patch cadence](https://captechgroup.com/services/managed-it-solutions) set around staffing and change-control windows now carries a wider gap against how quickly a disclosed flaw can be used against you.

Your **internet-facing assets** carry most of this exposure. Public websites, web applications, and customer portals are the systems automated reconnaissance and vulnerability scanning can reach first. The sectors GTIG ties to observed activity reflect this:

- **Media and entertainment** websites in France
- **Education** organizations
- The **aerospace** sector in the US
- Organizations across **Asia** subjected to systematic scanning

Many of these organizations run large public web footprints with limited security staff, which is the profile automated exploitation reaches most easily.

The attacker population is also changing. Hartman noted that AI lowers the cost and expertise needed to find flaws, exploit them, and move through networks, which puts those capabilities within reach of smaller, less sophisticated groups. GTIG assesses with high confidence that cybercrime actors less concerned with stealth will keep adding autonomous tools, because wider reach offsets lower success rates. If your risk model assumes only well-funded adversaries would target your environment, that assumption no longer holds. GTIG also expects intrusions to become nearly fully agentic, with less human involvement and more stealth as the underlying models improve.

These shifts carry into **cyber insurance**. If your application or renewal asks how quickly you remediate critical vulnerabilities on exposed systems, those answers describe your exposure window directly. After an incident involving a flaw that had a published patch for weeks, insurers may look closely at what you attested about remediation timelines.

**Vendor risk** follows the same logic. Your hosting providers, SaaS platforms, and managed service partners face the same compressed timeline, and a contract that gives them a month to apply critical fixes leaves your data exposed for that month. A separate cost line appears through LLMJacking, where attackers steal credentials for AI services and run up usage on the victim's account. GTIG cited one case where cumulative abuse reached the earlier total. Any API keys your developers or vendors hold for LLM services are now a direct billing risk.

Board reporting is affected as well. Vulnerability metrics usually reach directors as the share of patches closed within an agreed SLA. GTIG's findings make a different figure more relevant to your board: how many days internet-facing systems remained exploitable after the vendor published a fix, since that interval is what automated tools target. Hartman described the report as an early warning and said organizations should assume these capabilities will spread to more nation-state and criminal operators.

## Shortening Exposure Windows and Detecting AI-Assisted Exploitation

The most important step you can take is to list every internet-facing web application you run and patch the ones with known-exploited flaws ahead of your normal cycle. The frameworks GTIG describes go after code that has a public fix but has not yet been deployed, so this list is your highest-priority patch queue. The steps below are grouped by the functions of the NIST Cybersecurity Framework and ordered by urgency within each function.

### Identify

- **Inventory exposed systems:** Have your IT lead or provider list every public website, content management system, customer portal, and remote access gateway, including sites hosted by marketing agencies or other third parties.
- **Cross-check against CISA's Known Exploited Vulnerabilities (KEV) catalog:** Any match on an internet-facing system goes to the top of the queue.
- **Inventory AI credentials:** List the LLM API keys and cloud AI accounts your staff and developers use. GTIG reports that LLMJacking actors hunt for these keys and resell access through illicit LLM Mixer marketplaces.

### Protect

- **Run a separate fast patch track for edge systems:** Patch KEV-listed and public-facing flaws as soon as fixes are tested, outside the monthly cycle Hartman says many operators struggle to meet.
- **Use virtual patching when a fix can't ship:** Place a web application firewall (WAF) rule in front of the vulnerable component. A virtual patch blocks the known exploit path until the real update is installed.
- **Lock down AI keys:** Store keys in a secrets manager and keep them out of code repositories. Scope each key to one project and set billing alerts so unexpected usage gets flagged.
- **Govern internal agentic tools:** Keep an approved list of AI agent frameworks your teams may run. Require review before any agent executes code on production hosts or holds standing credentials.

### Detect

GTIG notes that heavy LLM use creates forensic noise, and it expects cybercrime groups to accept noisier operations in exchange for reach. That noise gives your monitoring something to catch.

- **Machine-speed request patterns:** Watch web server and WAF logs for bursts of probing across many endpoints, rapid movement from reconnaissance to exploit attempts, and requests aimed at a component soon after its vendor advisory appears.
- **Webshells:** The exploit\_agent server held webshell deployment scripts. Alert on new or modified script files in web directories and on web server processes launching command shells.
- **Tool artifacts:** Load any indicators your threat intelligence feeds publish for HexStrike AI, PentAGI, Cairn, or ZeroHollow. The GTIG summary does not list specific indicators, so rely on your vendor's feeds.
- **AI billing spikes:** Treat sudden usage jumps on your LLM accounts as a possible stolen key.

### Respond

Hartman's advice is to move toward automated defenses that detect and respond at machine speed. In practice, pre-approve your provider to isolate a web server that shows signs of a webshell without waiting for a meeting. After isolation, preserve logs, look for staged data (the exploit\_agent server held stolen victim data), and rotate any credentials stored on the affected host.

### Recover

- **Rebuild, then patch:** Restore compromised web servers from known-good images and apply the fix before returning them to service.
- **Adopt continuous attack surface management:** Replace the yearly inventory with ongoing discovery of new public assets.
- **Test with AI on your own assets:** Use AI-assisted defensive testing to check your own patch backlog for exploitable flaws before outsiders do.
- **Track one metric:** Record the days between each vendor advisory and your deployment, and review that number every quarter.

If you are unsure how quickly your public websites and portals get patched, the provider or agency that hosts them can give you their current patch timeline and WAF coverage. A [security assessment](https://captechgroup.com/security-assessment) is a sensible next step: a 15-minute call followed by written findings on five areas, including security controls and remote access.

<!-- AI:SCHEMA: Schema.org description of canonical page in JSON-LD format -->
<!-- AI:SCHEMA:BEGIN format=jsonld scope=page -->

```json
{
    "@context": "http:\/\/schema.org",
    "@graph": [
        {
            "@type": "Article",
            "author": {
                "@id": "https:\/\/captechgroup.com\/#brian_0fd5dfcdbc"
            },
            "dateModified": "2026-11-18T12:34:31Z",
            "datePublished": "2026-11-18T12:34:31Z",
            "description": "Google's GTIG found exploit_agent, an AI framework that diffs patches and builds n-day exploits without human input, shrinking the window to patch web apps.",
            "headline": "AI Framework Autonomously Weaponizes Flaws in Real Attacks, Google Reports",
            "image": [
                {
                    "@type": "ImageObject",
                    "url": "https:\/\/images.captechgroup.com\/cdn-cgi\/image\/width=1200,format=webp,quality=85\/threat-intel\/cc0d7c15f0.jpg",
                    "caption": "",
                    "description": "AI exploit_agent framework scanning web servers at machine speed, hitting one unpatched server in n-day exploitation",
                    "width": 1200,
                    "height": 685
                }
            ],
            "inLanguage": "en-GB",
            "mainEntityOfPage": {
                "@type": "WebPage",
                "url": "https:\/\/captechgroup.com\/threat-intelligence-center\/ai-framework-autonomously-weaponizes-flaws-in-real-9cebe9"
            },
            "publisher": {
                "@id": "https:\/\/captechgroup.com\/#defaultPublisher"
            },
            "url": "https:\/\/captechgroup.com\/threat-intelligence-center\/ai-framework-autonomously-weaponizes-flaws-in-real-9cebe9"
        },
        {
            "@type": "Person",
            "name": "Brian",
            "@id": "https:\/\/captechgroup.com\/#brian_0fd5dfcdbc"
        },
        {
            "@id": "https:\/\/captechgroup.com\/#defaultPublisher",
            "@type": "Organization",
            "url": "https:\/\/captechgroup.com\/",
            "logo": {
                "@id": "https:\/\/captechgroup.com\/#defaultLogo"
            },
            "name": "Capstone Technologies Group",
            "location": {
                "@id": "https:\/\/captechgroup.com\/#defaultPlace"
            }
        },
        {
            "@id": "https:\/\/captechgroup.com\/#defaultLogo",
            "@type": "ImageObject",
            "url": "https:\/\/captechgroup.com\/images\/hotlink-ok\/logo-light.jpg",
            "width": 1300,
            "height": 300
        },
        {
            "@id": "https:\/\/captechgroup.com\/#defaultPlace",
            "@type": "Place",
            "address": {
                "@id": "https:\/\/captechgroup.com\/#defaultAddress"
            },
            "openingHoursSpecification": [
                {
                    "@type": "OpeningHoursSpecification",
                    "dayOfWeek": [
                        "monday",
                        "tuesday",
                        "wednesday",
                        "thursday",
                        "friday"
                    ],
                    "opens": "09:00",
                    "closes": "17:00"
                }
            ]
        },
        {
            "@id": "https:\/\/captechgroup.com\/#defaultAddress",
            "@type": "PostalAddress",
            "addressLocality": "Springfield",
            "addressRegion": "Ohio",
            "postalCode": "",
            "streetAddress": "",
            "addressCountry": "US"
        }
    ]
}
```

<!-- AI:SCHEMA:END -->

