cd /news/ai-safety/google-spots-the-first-hacker-ai-sys… · home › topics › ai-safety › article
[ARTICLE · art-171168] src=csoonline.com ↗ pub= topic=ai-safety verified=true sentiment=↓ negative

Google spots the first hacker AI system that finds and exploits flaws autonomously

Google Threat Intelligence Group and Mandiant reported the first observed hacker-operated AI framework that autonomously discovers software vulnerabilities and executes exploits, found on an exposed staging server and identified as a multi-agent framework called exploit_agent. The framework connects to multiple large language models and is configured to examine code changes after security patches, identify flaws, and generate and run exploits without an operator directing each step; Google found data exfiltrated from targeted victims on the server, with evidence the operator focused on French media, entertainment and education websites. Google also cited a separate campaign in which a threat actor likely used AI tools to develop and test exploits against PaperCut print management server vulnerabilities, hitting hundreds of organizations four days after disclosure, while Omdia senior analyst Jonathan Ong said the activity is "more of an acceleration of existing attack techniques.

read4 min views1 publishedNov 17, 2026

Google said it had identified a hacker-operated AI framework designed to discover software vulnerabilities and execute exploits without human intervention, marking the first time its researchers had observed such a system being used by threat actors.

The findings, published in a report by Google Threat Intelligence Group (GTIG) and Mandiant, suggest that attackers are moving beyond using generative AI to assist individual tasks and are beginning to automate larger portions of cyberattacks.

The system could give attackers a way to exploit publicly disclosed vulnerabilities, known as n-days, before companies finish applying security updates. Google’s researchers discovered it on an exposed staging server and identified it as a multi-agent framework called exploit_agent.

The framework connects to multiple large language models (LLMs) and is configured to examine changes in application code following security patches, identify vulnerabilities, and generate exploits without requiring an operator to direct each step. It can also execute the exploits automatically.

Google found data exfiltrated from targeted victims on the server, along with scripts designed to support further activity after an initial compromise. Evidence suggested that the operator was particularly interested in French websites associated with the media and entertainment industries and the education sector. The report did not establish whether the framework’s autonomous operations were responsible for the theft.

Google also cited a separate campaign involving vulnerabilities in PaperCut print management servers. According to research by GreyNoise referenced in Google’s report, a threat actor likely used AI tools to develop and test exploits before launching a campaign against hundreds of organizations four days after the vulnerabilities were disclosed.

In separate cases, Google observed other AI frameworks being used by attackers. CAIRN was deployed in intrusion activity targeting organizations that included the US aerospace sector, while a customized platform based on PentAGI was used to conduct reconnaissance and vulnerability scanning against organizations across Asia.

Despite the growing use of AI in cyberattacks, the underlying techniques remain largely unchanged, according to Jonathan Ong, senior analyst at Omdia.

“This is more of an acceleration of existing attack techniques,” Ong said. “These are pentesting/red-teaming frameworks with established techniques, albeit significantly sped up with an LLM model.”

Keith Prabhu, CEO of Confidis, said there is still limited evidence that AI makes individual attacks more successful. The greater risk, he said, is that attackers could launch more attempts in less time, potentially increasing the number of successful compromises even if each attempt is no more effective.

Ong said the frameworks typically generate considerable detectable activity, meaning enterprises with properly configured and monitored endpoint detection and response (EDR), extended detection and response (XDR), or managed detection and response (MDR) capabilities should still be able to identify much of the malicious activity.

Google also warned that although the frameworks generate activity that makes detection easier, they remain capable of achieving their objectives.

Rather than overhauling their defenses, security leaders should continue prioritizing established security practices, particularly cloud and identity hardening. Ong pointed to multifactor authentication, strong passwords, and measures to block repeated login attempts as effective safeguards.

“The one thing that has changed is that the containment and response timeline needs to be faster, to match the speed at which attackers using these frameworks are moving,” Ong said.

Prabhu said the problem is compounded by fragmented security visibility and response processes that still depend on manual decisions. He recommended that CISOs prioritize continuous asset discovery and risk-based vulnerability remediation. He also called for better detection capabilities that connect suspicious activity across enterprise environments.

Amit Kumar Jena, development manager for AI at Kanerika, said organizations must give particular attention to vulnerabilities in systems exposed to the internet, where attackers could exploit flaws soon after they become public.

“Internet-facing systems should be treated as their own class, with a fast-track patch path and virtual patching while fixes are tested,” Jena said.

For incident containment, Prabhu suggested automating actions that are well understood and can be reversed, while retaining human approval for decisions that could have significant operational consequences. Jena also recommended defining in advance which containment actions could be automated, citing endpoint isolation and session revocation as examples.

“The aim is not to remove human oversight, but to eliminate avoidable delays and regularly test that response works at the pace of an attack,” Prabhu said.

Open-source agentic frameworks could lower the technical barriers for less-experienced cybercriminals by helping them coordinate existing tools and adapt their attacks based on results, Prabhu said. However, these capabilities do not eliminate the need for technical expertise or guarantee successful compromises.

Prabhu said smaller organizations and service providers with limited security resources could be particularly exposed.

The risk is not necessarily concentrated in any particular industry, he added. Organizations with poorly maintained legacy systems or weak identity controls could face greater exposure regardless of their sector.

── more in #ai-safety 4 stories · sorted by recency
── more on @google 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/google-spots-the-fir…] indexed:0 read:4min 2026-11-17 · —