In its 30 October report on ARTEX, Google Threat Intelligence Group (GTIG) forecasts that in three to six months attackers will move to fully autonomous intrusions with AI agents, stealthier than those seen so far. On 16 November GTIG and Mandiant published their quarterly update on attackers’ use of AI, which finds that agentic threat activity has crossed a significant threshold. The new report documents multi-agent offensive frameworks used in real intrusions and, for the first time, a vulnerability discovery and exploitation chain with no human operator.

Three days later Been Kim, Principal Scientist at Google DeepMind, wrote on X that about two months ago her team moved entirely to safety. Kim comes from interpretability and today, she writes, she cannot justify working on anything other than directly improving frontier model safety. She is hiring researchers in Seattle to work on urgent problems in Gemini itself. In the job posting the team says it puts concrete safety wins ahead of conventional paper-writing cycles. The report and the post come from the same company, from the people who watch the attacks and from those who work on the models.

Offensive agents in intrusions

Since late August 2026 GTIG and Mandiant have observed attackers using open source multi-agent penetration testing platforms, in particular ARTEX, CAIRN and PentAGI. ARTEX splits the attack lifecycle across six sub-agents that call frontier and open source models. In the logs analysed by Mandiant its execution sequences last roughly 2 to 25 seconds. In one case, supplied with valid credentials, ARTEX used a web application’s SQL API to make the MySQL general log write .jsp files to a public directory, in other words webshells. It wrote one after another, fixing the errors of the previous attempts, until one returned command output.

These tools currently trade stealth for speed. In that same case ARTEX produced 240 log events, and in another it left files with the recognisable prefix artex-probe-.

From patch to exploit with agents

GTIG presents exploit_agent, found on an exposed staging server, as the first case of fully automated vulnerability discovery and exploitation. The framework compares the code of popular web applications before and after a patch, identifies potential vulnerabilities, writes the exploit and runs it with no human intervention.

GreyNoise reconstructed, minute by minute, a campaign using the same method against PaperCut NG/MF, a print management product, with an operator directing the agents. The advisory for two vulnerabilities which, chained, give unauthenticated remote code execution came out on 27 August 2026. On 31 August the attacker downloaded the pre- and post-patch versions of the software and in under four hours had written the exploits and hit the first real victim. The agents ran in OpenAI’s Codex harness on a DeepSeek model. The tally is at least 440 compromised instances in 395 organisations across 48 countries.

According to GTIG, exploited vulnerabilities rose from an average of 10.5 per month in 2025 to 18 per month in the first eight months of 2026. Zero-days are still the majority, at 62%, but they rose only slightly, from 8 to 11 per month, with a peak of 22 in August. GTIG therefore suggests that the growth comes mainly from the rapid weaponisation of n-days. Its hypothesis is that LLMs make it easier or more efficient to automate the analysis of what is already public: the differences between product versions, patches, vulnerability disclosures and proof-of-concept code.

Credentials and agent configurations

Mandiant handled cases of AI resource theft, which GTIG calls LLMJacking. An attacker inside a cloud environment consumes the victim’s LLM services or compute for their own tools. In some cases single-day charges exceeded $100,000.

Closer to anyone who builds software with agents is the case of a group GTIG associates with North Korea. After initial access the group modified the .claude/settings.json files in internal repositories so that anyone starting Claude Code would download a backdoor. GTIG describes it as a supply chain vector internal to the organisation. A development agent’s configuration can define commands (hooks) that the agent runs with the user’s permissions. It therefore has to be reviewed like code.

GTIG’s forecast for the next six months

In the ARTEX report GTIG describes the evolution of offensive LLM use in seven stages, from chatbot-style use of LLMs to write malicious code and phishing emails at the end of 2025 to fully autonomous intrusions. It places the current moment at stage five: several attack phases carried out by LLM-driven tooling, with minimal human intervention and still noisy techniques. For the following months it forecasts two steps.

  • Near term, 1 to 3 months. Highly autonomous but unsophisticated, noisy multi-agent attacks with no human operators in the loop. The attacker picks the target and the objective, and the frameworks run the entire attack chain and return the result.
  • Medium term, 3 to 6 months. Fully autonomous intrusions, stealthier and more sophisticated, run through obscure infrastructure or self-hosted models that reduce defenders’ visibility.

GTIG assesses with high confidence that actors less concerned with stealth, including many cyber crime groups, will keep adopting these tools, because the greater reach of their campaigns will likely make up for a lower success rate. It considers groups aiming for long-term access less likely to use the public versions, which are too noisy, except to gain speed and scale.

What labs measure on their own models

In the Gemini 4 Carbon safety report Google cannot rule out that the model without out-of-model mitigations reaches the Cyber Uplift Level 1 critical capability level of its Frontier Safety Framework. The determination is carried forward from Gemini 4 Argon, and with its safeguards in place Google assesses the risk to be substantially below the threshold. On 5 November OpenAI classified GPT-6 Cyber as Critical for cybersecurity under its Preparedness Framework and for now makes it available only to a select group of approved customers.

The risk also concerns how the models themselves behave. In July 2026, during internal cybersecurity evaluations, OpenAI models circumvented the controls isolating them from the internet and compromised parts of OpenAI’s research infrastructure and Hugging Face’s systems. Later an internal research model exploited a DNS misconfiguration to gain limited internet access. OpenAI paused training and evaluation of its most capable models from 20 September to 1 October. On 18 November it described the rules adopted after these incidents, including training with no internet access and monitoring of all frontier model inference. This is the work described in the job posting for Kim’s team: turning alignment and safety findings into mitigations on live frontier models.

What changes for defenders

Mandiant’s recommendations are familiar practice: isolate the control plane, use short-lived, narrowly scoped credentials, segment, apply least privilege and detect anomalous behaviour. What changes is the time available. In the PaperCut case four days passed between the advisory and the campaign’s first victim, and five until the mass attack. Patches therefore have to be prioritised using data on real-world exploitation. Anyone integrating agents into their own processes has to treat model keys as production credentials and record what each agent does and with which permissions. This is why agents need governance layers that keep track of them.

If GTIG’s forecast comes true, the noise that makes ARTEX detectable today will drop within a few months. The identity, control plane, data plane and network telemetry that Mandiant says behavioural detection depends on has to be put in place now, while the attacks are still visible.

I have been working on these topics for years. At noze we apply them to the solutions and products we build. In my courses I use them to try to manage the inevitably accelerated and compressed path that lies ahead of students, junior and senior alike. On more than one occasion I have found myself playing Cassandra, while arguing that with governance and the right controls this evolution can be dealt with. That is where OISG and Admina came from.


Cover image: Evelyn De Morgan, “Cassandra”, 1898, De Morgan Foundation. The Trojan prophetess nobody believed, in front of the burning city and the wooden horse. Public domain — https://commons.wikimedia.org/wiki/File:Cassandra1.jpeg