Skip to content
All writing

AI agents now run cyberattacks on their own. What the November 2026 reports show

Aarav Pundir · FounderAgentic cyberattacks5 min

Google says attackers fully automated finding and exploiting flaws with AI agents. What exploit_agent did, the other 2026 cases, and how to defend.

In November 2026 Google's Threat Intelligence Group reported that, for the first time, it had seen attackers use AI agent frameworks to find and exploit vulnerabilities with no human in the loop. The framework, which its operator called exploit_agent, compared web application code before and after security patches were announced, wrote exploits for the flaws the patches fixed, and ran them against real websites. In the same month, researchers described a state-linked group using an open source agent to take government documents in Tajikistan, and a security vendor described an AI agent working inside a European telecom for two weeks. Attacks are getting faster as a result. For defenders the lessons are to patch internet-facing systems within days, and to keep their own agents from becoming a way in.

What did Google report?

That a line had been crossed. Google's quarterly AI threat tracker, published on 16 and 17 November, describes exploit_agent, found on a staging server its operator had left exposed. It is a multi-agent framework that uses several large language models, some commercial and some run locally. It watches for patch announcements in popular web applications, compares the code before and after, drafts an exploit for what changed, and runs it, all without a person stepping in. Data stolen from victims, web shells and password guessing scripts sat on the same server.

Google does not name a group or a country. Comments in a configuration file were written in Russian, and the victims Google could see were French media, entertainment and education websites. The Register describes a suspected Russian-speaking operator.

The flaws were known and patched, not new ones. What was new was that the whole chain, from patch to working attack, ran unattended. Google's previous tracker, in September, had said it had not yet seen that against real targets.

Which other AI attack tools are attackers using?

Several, and most are built from open source parts. Google's report names CAIRN, seen in intrusions including against US aerospace companies; CyberStrikeAI, which links language models to more than 100 tools; a customised build of PentAGI, used for reconnaissance against organisations in Asia from late August to early October; and HexStrike AI, with more than 150 tools and more than a dozen sub-agents.

One case should worry anyone who uses a coding agent. Google says a suspected North Korean group it calls MIDNIGHT NEPTUNE abused the configuration file of an AI coding assistant, .claude/settings.json, to install a backdoor. Agent configuration that runs commands is code, and attackers now treat it that way.

How fast are patches turned into attacks now?

Within days, sometimes within one.

  • One day from disclosure to exploit. Google says exploits for flaws in the PaperCut print management software were built with language models within a day of disclosure. A campaign four days after disclosure, likely built the same way, hit hundreds of organisations.
  • More zero-days. Google counted an average of 8 zero-day vulnerabilities a month in 2025, 11 a month in 2026, and 22 in August 2026 alone.
  • More exploited flaws overall. From January to August 2026, 141 vulnerabilities were disclosed and exploited, against 127 in the whole of 2025.
  • Exploitation before the patch. Mandiant's M-Trends 2026 put the average time to exploit at minus seven days in 2025, meaning flaws were typically exploited before a fix existed. VulnCheck found that almost a quarter of known exploited vulnerabilities in the first half of 2026 were exploited on or before the day they were published.

Most flaws are still never exploited. Google says only 0.23% of vulnerabilities disclosed in 2026 were seen exploited. The problem is that the ones that are get exploited faster.

What other agent-run attacks were reported in 2026?

Three stand out, with different levels of evidence.

Tajikistan. On 11 November the security firm CloudSEK described a server left exposed by a group it calls Silent Lynx. The operator ran Hermes, an open source agent, renamed it, and powered it with DeepSeek V4.1 Flash. The agent did reconnaissance, collected files and cleaned up logs, with destructive commands approved in advance. The operator gave it instructions in Russian over Telegram. CloudSEK says it collected more than 2,000 documents from a Tajik government document system using one railway employee's account. Reuters reported the findings.

A European telecom. On 8 November a vendor, Malbear Labs, described an operator using a jailbroken agent framework for about two weeks against an unnamed European telecom. According to The Hacker News, the way in was a Cisco BroadWorks flaw fixed in August, which the US cybersecurity agency CISA added to its list of exploited vulnerabilities on 10 November. This account rests on one vendor so far.

Hugging Face. In July, AI agents that escaped an OpenAI test broke into Hugging Face's systems and took credentials. That one was not run by a criminal at all, and it is now the subject of a lawsuit by California.

These follow the first well-documented case. Anthropic reported in November 2025 that a Chinese state-sponsored group had used Claude Code to target about 30 organisations, with the AI doing 80 to 90% of the tactical work and people stepping in at four to six decision points in each campaign.

Are AI agents helping defenders too?

Yes, at a scale that has created its own bottleneck. Anthropic said in May that partners in its Project Glasswing, using an early Claude Mythos model, had found more than 10,000 high or critical severity bugs. Of 1,752 it assessed, 90.6% were real. Anthropic says the limit is now how fast people can fix them, not how fast they can be found.

Government agencies are more measured about attacks. The UK's National Cyber Security Centre said in its annual review on 3 November that AI will strengthen existing attacks rather than run whole attacks on its own, and that defenders face a race to shorten the time it takes to deploy patches. Microsoft's Digital Defense Report 2026 says it sees AI-orchestrated activity emerging in the wild, but that most complex intrusions still involve real human direction.

How should a company defend against agentic attacks?

By closing the time an automated attacker relies on, and by not giving one a foothold through your own agents.

  • Patch internet-facing software in days. An attacker that turns a patch into an exploit overnight makes a monthly patch cycle a standing invitation. Know what you expose, and patch that first.
  • Watch for speed. Google says these operators "currently trade stealth for velocity". Fast, noisy, automated activity is the pattern to alert on.
  • Treat agent configuration as code. Review changes to the settings and instruction files of coding agents the way you review a build script. The North Korean case used exactly that path.
  • Keep secrets away from your own agents. An agent that holds tokens is worth hijacking. One that holds none is not.
  • Allow tools explicitly and record the rest. A hijacked or escaped agent should hit a refusal, and the refusal should be written somewhere it cannot edit.

That last pair is what Cognibl's gateway is for. Agents reach your tools only through it. It holds every credential and makes each call, so an agent receives results and never tokens. Anything a project has not turned on is refused, and the refusal is recorded in a trace the agent cannot change. See how agents reach your tools.