Google’s threat intelligence team says attackers’ use of agentic AI has crossed a significant threshold. Here it means AI that carries out multi-step tasks with little human direction, not just a chatbot that answers questions.
Key takeaways
- Google Threat Intelligence Group (GTIG) has published an AI Threat Tracker update saying agentic threat activity has crossed a significant threshold.
- Agentic AI means systems that plan and carry out a series of actions with limited human supervision, unlike tools that only generate text when asked.
- In cyber attacks, the worry is that agents could automate slow, skilled work such as reconnaissance, writing code and adapting to defences.
- The headline does not give the exact scale, the groups involved or the techniques observed, so readers should not assume details beyond it.
- Defenders face pressure to automate their own detection and response so they can keep pace with attacks that may move faster than before.
What is happening with AI agents and cyber attacks?
Google Threat Intelligence Group, the security research unit within Google Cloud, has released an update to its AI Threat Tracker titled “Agentic Threat Activity Crosses a Significant Threshold”. The title says malicious use of agentic AI has reached a level the researchers consider meaningful, rather than staying experimental or marginal.
What exactly that threshold is cannot be confirmed from the title alone. It could be a rise in how often such activity is seen, a change in how capable it is, or evidence that agents are being used in real operations rather than in tests. This article does not know which threat actors were named, which sectors were targeted or what figures the report contains. It focuses on what agentic threat activity means and why a finding like this matters.
Why is this in the news now?
Many organisations now use AI agents: software that receives a goal, breaks it into steps, calls tools, checks the results and carries on until the task is done. As these products have spread to coding, customer service and IT administration, security researchers have watched for attackers adopting the same approach.
Reports from large threat intelligence teams carry weight because those teams see a lot of malicious activity on their platforms and services. A statement from such a team that a threshold has been crossed suggests a change in kind, not a gradual drift. That is why it has caught attention well beyond the security industry. The exact timing and the evidence behind the conclusion are set out in the original publication, not in this summary.
What background does a newcomer need?
The phrase “AI in cyber attacks” covers several quite different things, and it helps to separate them.
The first is AI as an assistant. Attackers use chatbots much as other people do: to translate text, write more convincing phishing emails, fix code or summarise technical documents. This makes existing work faster but does not change what attackers can do in principle.
The second is AI built into malicious tools. Here a piece of malicious software contacts a language model while it is running, for example to create new commands or change its own code. This can make the tool harder to detect with fixed signatures.
The third is agentic AI, which is what this report concerns. An agent can be given a broad goal and left to pursue it through many steps. It might scan for exposed systems, rank weaknesses, try to exploit them, react to errors and report back. A human operator supervises much less closely than with a chatbot.
Moving from the first category to the third is significant because it shifts AI from helping an attacker to doing part of the attacker’s job. Many attacks have been limited by skilled human labour. Anything that automates that labour could change how many attacks happen and how quickly they unfold.
Who is affected, and how?
The effects fall unevenly across several groups.
Organisations with internet-facing systems are the most obvious exposure. If agents make reconnaissance and first attempts at intrusion cheaper, unpatched servers, misconfigured cloud services and weak login portals may be found and probed more often.
Smaller organisations may be hit relatively hard. They have often relied, without saying so, on not being worth a skilled attacker’s time. If automation lowers the cost of targeting them, that informal protection weakens.
AI providers are affected because attackers may be using their models and agent platforms. Providers have to tell legitimate automation apart from abuse, which is hard when the same capability serves both.
Security teams must deal with a possibly faster tempo. An agent working around the clock can shorten the time between first contact and real damage.
Individuals feel the effects mostly indirectly: through breaches of services they use, more persuasive scams, and pressure on the organisations that hold their data.
Where do informed people disagree?
Security researchers do not fully agree on how much AI changes the threat landscape, and a report like this does not settle the question.
One view holds that the effect has so far been gradual. On this view, AI mainly speeds up work attackers could already do, and the basic weaknesses are unchanged: unpatched software, stolen credentials and social engineering. If so, the usual defensive advice still covers most of the risk.
Another view treats agentic capability as a turning point. If agents can reliably chain steps together without close supervision, a small team could run far more operations, and less skilled actors could attempt intrusions that once needed expertise.
There is also debate about measurement. Activity seen on one provider’s platform may not reflect what happens elsewhere, including on openly available models that run outside any provider’s oversight. Some observers also warn that vendor threat reports, while useful, come from organisations that sell security products. Others reply that these organisations are among the few with enough visibility to see the trends at all.
How far this particular report’s evidence supports the stronger reading cannot be judged from its title.
What are the practical implications?
For defenders, most of the response builds on existing good practice rather than replacing it, though with more urgency.
- Reduce the exposed surface. Automated reconnaissance rewards attackers who find forgotten servers and open services. An accurate inventory of internet-facing assets matters more when scanning is cheap.
- Patch faster. If agents shorten the gap between a vulnerability being disclosed and being exploited, slow patch cycles become riskier.
- Strengthen identity controls. Multi-factor authentication that resists phishing, careful handling of credentials and the principle of least privilege limit what any intruder can do, whether human or automated.
- Automate detection and response. Teams that rely entirely on manual triage may struggle to match the speed of automated attacks. Many organisations are looking at their own AI-assisted tools for this reason.
- Secure your own agents. Organisations deploying AI agents should treat them as privileged software: limit their permissions, log their actions, and watch for manipulation through malicious instructions hidden in the content they read.
For AI developers, the implication is closer monitoring of how agent features are used and clearer policies for cutting off abuse.
What should readers watch next?
Several developments will show whether the threshold described becomes a lasting shift:
- Whether other threat intelligence teams and government cyber agencies publish similar observations, which would support or qualify Google’s findings.
- Whether confirmed incidents emerge in which agentic AI clearly played a central role, as opposed to a supporting one.
- How AI providers change their safeguards, terms of use and abuse reporting in response.
- Whether regulators and standards bodies start addressing the security of agentic systems specifically.
- How defensive tools change, especially products that use agents to investigate alerts and contain intrusions automatically.
Frequently asked questions
What is agentic AI in cybersecurity?
Agentic AI refers to systems that receive a goal and pursue it through a series of actions, such as running tools, reviewing results and choosing next steps, with limited human supervision. In cybersecurity the term covers both defenders using agents to investigate alerts and attackers using them to automate parts of an intrusion. The defining feature is autonomy across several steps, not simply producing text when prompted.
What did Google’s AI Threat Tracker report say?
Google Threat Intelligence Group published an AI Threat Tracker update saying that agentic threat activity has crossed a significant threshold, meaning malicious use of AI agents has reached a level its researchers consider meaningful. The specific evidence, threat actors and figures behind that conclusion are not covered in this summary. Readers wanting those details should consult the original publication from Google Cloud.
Are hackers using AI to attack companies?
Security researchers have reported that threat actors use AI tools for tasks such as writing phishing messages, researching targets and producing code. The latest Google update suggests use is moving beyond simple assistance towards more autonomous, agent-driven activity. How widespread this is, and how much harm it has caused so far, is not established in the material available here.
How is an AI agent different from a chatbot?
A chatbot answers one prompt at a time and waits for the next instruction. An agent is given a broader objective and works towards it by itself, calling tools, checking outcomes and deciding what to do next. For attackers, this difference matters because an agent could handle a sequence of tasks that would otherwise need continuous human effort and expertise.
How can organisations defend against AI-driven attacks?
The core defences are familiar: keep an accurate inventory of exposed systems, patch quickly, use phishing-resistant multi-factor authentication and limit account privileges. What changes is urgency, since automated attacks may move faster. Many organisations are also adopting automated detection and response, and securing their own AI agents by restricting permissions and logging their actions.
Does AI make cyber attacks more dangerous for small businesses?
Possibly. Small businesses have often been ignored because targeting them took skilled time for little reward. If AI agents make reconnaissance and first intrusion attempts cheaper, more small organisations could be probed. Basic measures, including updates, strong authentication, backups and staff awareness, remain the most effective protection, and many are inexpensive to put in place.
Sources and further reading
- Google Threat Intelligence Group, via the Google Cloud blog: the AI Threat Tracker update on agentic threat activity.
- National cybersecurity agencies: general guidance on patching, identity security and incident response.
- AI developers’ published usage policies and safety reports: how providers describe abuse monitoring for agent features.
- Independent security research publications: wider analysis of how generative and agentic AI is being used by threat actors.
Surfaced from the google:US signal “AI-driven cyber attacks”. AI-assisted draft, editorially reviewed.

