Google — AI Agents Automate Exploits, Steal Credentials and Scale Cybercrime

Publication date: November 16, 2026
Category: AI attacks (LLM/LocalAI)

Introduction

The cybersecurity landscape has undergone a profound operational shift: malicious actors have moved beyond manual chatbot prompting to deploy autonomous multi-agent frameworks and large language models (LLMs). According to a recent report by Google’s Threat Intelligence Group (GTIG), these frameworks—such as ARTEX, CAIRN, CyberStrikeAI, and PentAGI—allow attackers to perform differential patch analysis, automate reverse engineering, and execute multi-stage intrusions at unprecedented operational speeds without direct human intervention. This phenomenon dramatically accelerates the exploitation of known vulnerabilities (n-day) and drives massive financial abuse through LLMjacking.

What is the Autonomous Agent Phenomenon and Involved Vulnerabilities? (General Analysis)

AI agent-based automation represents the convergence of traditional offensive security tools with the cognitive capabilities of language models. These agents possess the ability to discover simple vulnerabilities, synthesize customized payloads via LLMs, and adapt their strategy in real time based on feedback from the target environment.

Concurrently, this threat ecosystem directly interacts with critical infrastructure and software flaws. Key examples include recent vulnerabilities in the PHP engine (CVE-2026-109618 and CVE-2026-109619) and the LiteLLM platform (CVE-2026-40217).

Confirmed Vulnerabilities Summary (NVD Data)

CVE Category (CWE) Impact CVSS Vector (summary)
CVE-2026-109618 CWE-416, CWE-755, CWE-824, CWE-1325 Remote Code Execution (RCE) 9.2 (Critical, CVSS v4.0) AV:N/AC:H/AT:P/PR:N/UI:N/VC:H/VI:H/VA:H
CVE-2026-109619 CWE-416, CWE-755, CWE-824, CWE-1325 Remote Code Execution (RCE) 9.2 (Critical, CVSS v4.0) AV:N/AC:H/AT:P/PR:N/UI:N/VC:H/VI:H/VA:H
CVE-2026-40217 CWE-420, CWE-94 Arbitrary Code Execution 8.8 (High, CVSS v3.1) AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H

Note: The scores and classifiers for the three CVEs above correspond to official confirmed records from the NVD database.

How Does It Work? (Technical Analysis)

The technical operation of these campaigns is split into two primary vectors: automated AI-driven attack orchestration and the exploitation of underlying software vulnerabilities.

  • Infection Flow and Automation: Frameworks like CAIRN and PentAGI utilize shared workspaces and connectors to over a hundred security tools. A specialized agent evaluates the environment, detects outdated components, and iteratively deploys exploits. In vulnerability research, experimental tools like exploit_agent autonomously perform differential analysis of web patches.
  • PHP Memory Flaws (CVE-2026-109618 and CVE-2026-109619):
    • CVE-2026-109618: PHP uses longjmp() for out-of-memory (OOM) errors to exit the current context. Certain contexts fail to properly account for this scenario, leaving request memory in an uninitialized state prior to allocation. Subsequent request cleanup procedures may read or branch on this memory, enabling strategies to achieve RCE.
    • CVE-2026-109619: Temporary global states pointing to request memory may not be correctly reset or may be repopulated after resetting, subsequently referencing already freed memory (Use-After-Free). Subsequent requests read or branch on this stale state before full initialization, facilitating code execution.
  • LiteLLM Privilege Escalation (CVE-2026-40217): This flaw allows authenticated remote attackers to execute arbitrary code via bytecode rewriting at the /guardrails/test_custom_code URI, easily turning limited-privilege keys into administrator credentials.
  • LLMjacking Infrastructure: Attackers monetize compromised resources via “LLM Mixers,” proxy networks distributing queries through stolen credentials and ephemeral service accounts to bypass rate limits, generating massive fraudulent cloud computing charges.

Affected Systems / Environments

  • PHP Environments: Web servers running PHP versions vulnerable to memory handling errors during OOM contexts and global state manipulation (both vulnerabilities listed in the CISA KEV catalog).
  • LiteLLM Platforms: LiteLLM deployments up to the version dated 2026-04-08 exposed to the custom code testing URI.
  • Cloud & AI Infrastructure: Cloud service providers, public source code repositories with leaked credentials, and language model APIs exposed without robust authentication.

Mitigation and Detection

Remediation

  • Software Patching: Immediately apply official security patches released by the PHP project to mitigate memory vulnerabilities (CVE-2026-109618 and CVE-2026-109619). Update LiteLLM to the vendor’s patched versions.
  • Strict Credential Management: Audit public repositories and development environments to prevent accidental leakage of API keys and session tokens. Implement mandatory credential rotation.
  • Access Control and Rate Limiting: Restrict access to administrative and testing endpoints in AI gateways, enforcing strict quota control policies and monitoring anomalous GPU resource consumption.

Detection

  • Behavioral Monitoring: Because AI frameworks generate distinguishable traffic artifacts and patterns when operating autonomously, prioritizing the supervision of anomalous API calls, mass creation, and deletion of ephemeral service accounts is critical.
  • Detection Rules: Deploy SIEM and EDR rules based on IOCs published by Google SecOps and Mandiant regarding frameworks such as ARTEX, CAIRN, and associated tooling.

“Attackers have moved past manual prompting; AI agent automation allows them to combine traditional techniques and scale malicious operations at a velocity that outpaces traditional human response capacity.”

Wrapping Up

The Google GTIG report highlights a critical turning point where artificial intelligence ceases to be merely a synthetic content vector and becomes the engine for autonomous cyberattack execution. The combination of automated frameworks, accelerated n-day vulnerability exploitation in critical components like PHP and LiteLLM, and monetization through LLMjacking requires organizations to modernize their defensive strategies, prioritizing behavioral visibility over static signature detection.

References