← News

Google Reveals Hackers Using Autonomous AI Frameworks to Exploit Software Patches

Google Threat Intelligence Group and Mandiant point out that threat actors are using multi‑agent AI frameworks, such as ARTEX and PentAGI, to automate reverse engineering of patches and generate exploits at machine speed.

By CyberX I.T SolutionsCibersegurança5 min readSource: Google Finds Hackers Using AI Agents to Build and Run Exploits From Software Patches ↗Ler em português →

Google disclosed, on November 16, 2026, that attack groups already use autonomous artificial intelligence (AI) frameworks to conduct multi-stage intrusions at machine speed.

Lead: The Google Threat Intelligence Group (GTIG), in partnership with Mandiant, identified the use of platforms such as ARTEX, CAIRN, CyberStrikeAI and PentAGI to automate patch reverse engineering, generate exploitation code, and orchestrate attacks without direct human intervention.

The cited frameworks are open-source projects that implement multi-agent architectures. Each agent performs specific tasks – information gathering, differential code analysis, payload generation – and shares results on a common “blackboard”, allowing the entire system to advance in a coordinated manner, as described in the GTIG report.

One of the most documented cases involves newly disclosed vulnerabilities in the PaperCut NG/MF print management software (CVE‑2026‑82078 and CVE‑2026‑81578). The analysis showed that AI agents were able to compare code versions before and after the patch, identify the flaw, and generate a functional exploit in minutes, drastically reducing the exposure window.

In addition to vulnerability exploitation, the report highlights the practice of “LLMjacking”, in which actors compromise language model credentials and cloud resources to generate fraudulent billing. Reported incidents point to charges exceeding US$ 100,000 in a single day, with some attacks accumulating more than US$ 1 million before interruption.

Another emerging trend is the integration of AI assistants into phishing platforms such as Jieyou, which allows users with little technical skill to configure phishing campaigns via natural language commands, expanding the reach of social engineering attacks.

While some reports, such as Google's own, claim that these attacks can occur “without any human intervention”, other sources point out that human oversight is still present in critical stages, such as credential delivery or result validation. This divergence reflects the complexity of measuring the real degree of agent autonomy.

For security teams, the rapid evolution of these frameworks requires adjustments to defense strategies: strengthening patch management processes, monitoring anomalous AI agent behaviors, and implementing safe LLM usage policies within organizations.

Technical Analysis CyberX

The ARTEX, CAIRN, CyberStrikeAI and PentAGI frameworks share an architecture based on multiple agents that operate in an OODA cycle (Observe, Orient, Decide, Act). Each agent can be specialized – for example, a metadata collection agent, a differential code analysis agent, and a third agent for payload generation using large language models (LLMs).

The patch reverse engineering mechanism starts with extracting two code versions: the vulnerable version and the fixed version. The analysis agent diff's the two sets, identifies security changes, and then feeds an LLM that proposes possible exploitation vectors. The LLM generates proof-of-concept code (PoC) that is automatically compiled and tested in controlled environments before being inserted into a deployment pipeline.

These tools can operate in cloud environments, leveraging provider APIs such as Google Cloud and Azure to scale the analysis. Automation reduces the “window of exposure” from days to minutes, especially when combined with “LLMjacking” services that allow the use of model credentials to generate large volumes of low-cost API calls.

From an impact perspective, automation increases the speed of compromise, but does not eliminate the need for execution privileges. Most agents require access to code repositories or build systems to obtain patch versions, which can be mitigated by strict access controls and monitoring of code reading activities.

Defensive recommendations include:

  • Apply patches immediately after disclosure, accompanied by artifact integrity verification.
  • Monitor patterns of access to code repositories and unusual API calls to AI services.
  • Implement LLM usage policies that restrict executable code generation to controlled environments.
  • Detect autonomous agent behaviors through audit event correlation (e.g., multiple code diff requests in short intervals).

CyberX SOC Intelligence Bulletin

The identified autonomous AI frameworks (ARTEX, CAIRN, CyberStrikeAI, PentAGI) are being used to automate patch analysis and exploit generation, drastically reducing the time between vulnerability disclosure and active exploitation.

Indicators of compromise (IoCs): So far, no hashes, IP addresses or specific domains associated with these frameworks have been disclosed in the consulted sources. It is recommended to monitor logs of code repository access and unusual LLM API calls.

Recommended mitigations:

  • Apply patches immediately and validate artifact integrity.
  • Restrict and audit access to versioning systems (Git, SVN) and AI APIs.
  • Implement anomalous behavior alerts in CI/CD pipelines, such as multiple sequential code diff executions.
  • Use autonomous agent detection tools based on resource usage patterns (e.g., spikes in LLM calls).

Detection Observations: The presence of autonomous agents can manifest as a large volume of code analysis requests in short periods, automatic generation of script files and communication between processes via temporary files or internal message queues. Correlating these signals with newly created API credential activities can indicate an ongoing attack.

…

Frequently Asked Questions

What autonomous AI frameworks were identified by Google and Mandiant?

Google Threat Intelligence Group and Mandiant identified open-source, multi-agent frameworks including ARTEX, CAIRN, CyberStrikeAI, and PentAGI being used by attack groups.

How do AI agents reverse-engineer software patches to generate exploits?

Analysis agents compare vulnerable and patched code versions, identify security changes, and feed the differences to an LLM, which proposes vectors and generates proof-of-concept code that is automatically compiled and tested.

What is LLMjacking as described in the report?

LLMjacking is the compromise of language model credentials and cloud resources to make fraudulent API calls, generating unauthorized billing that can exceed $100,000 in a single day.

Which vulnerabilities were targeted in the documented PaperCut NG/MF cases?

AI agents targeted newly disclosed PaperCut NG/MF print management software vulnerabilities CVE-2026-82078 and CVE-2026-81578 to create working exploits in minutes.

Esta análise foi útil?

Ainda não há avaliações.