Cybersecurity is heading toward a harder contest: defenders need to close weaknesses faster because attackers can use AI agents to keep looking for them after people go home.
Google’s Threat Intelligence Group said on November 16 that its latest AI Threat Tracker found bad actors using autonomous large-language-model frameworks to inspect software updates, infer what a patch fixed and develop an attack without continuous human involvement. That is a vendor’s threat assessment, not proof that every organization is already facing autonomous attacks.
The important change is the clock
Patch notes and software updates can reveal where a product was vulnerable. A capable automated system could repeatedly compare releases, test hypotheses and prepare a malicious path faster than a small security team can read every notice. The risk is not that AI makes every attack novel. It is that it can compress the time between a public fix and attempted exploitation.
That makes ordinary security discipline more valuable, not obsolete. Asset inventories, rapid patching, limited administrator access, tested backups and useful event logs are the controls that determine whether an alert becomes a contained incident or an expensive outage.
Do not let a security vendor write the whole conclusion
Google is also selling an AI Threat Defense offering that combines Gemini, Wiz, CodeMender and Mandiant capabilities. Its report is useful evidence of what its researchers say they observe, but a product announcement is not independent proof that one vendor’s system will solve the problem. Buyers should ask what was observed directly, what was simulated, and which parts of the response still depend on their own people and systems.
“Autonomous” also deserves precision. A tool can automate scanning, prioritization or code generation while a human still selects targets, supplies access or decides when to deploy the result. Inflating every automated step into a fully independent attacker makes security planning less clear.
What a practical response looks like
Start with the systems that can cause the most harm: internet-facing services, identity systems, remote-management tools and software with high-value data. Measure the time from a critical patch to deployment, and identify accounts or agents that can make changes without approval. Test whether an emergency credential revocation, rollback or service isolation actually works.
Security teams benefit when routine alert triage is faster. They lose if automation creates more opaque actions, wider permissions or a false sense that a dashboard has replaced incident response. The next useful evidence is independent reporting and technical detail showing how frequently these workflows succeed in real attacks and which controls stop them.
Frequently Asked Questions
Has Google proved that all cyberattacks are autonomous now?
No. Google reports observing autonomous LLM frameworks in attacker workflows. That does not establish that every attack is autonomous or equally capable.
Why can a patch create urgency?
A patch can indicate what weakness was fixed. Attackers may study the difference to identify systems that remain unpatched.
Should a small business buy an AI security product immediately?
First address known basics: supported software, multifactor authentication, backups, limited privileges and a workable patch process. Assess tools against a specific gap.
Can AI replace an incident-response team?
No evidence here supports that claim. Automation can help with repetitive work, but organizations still need accountable decisions and tested response procedures.
What is the most useful metric to track?
For critical systems, track the time from a relevant security fix to verified deployment, alongside the ability to detect and revoke risky access.