AI-assisted attacks are shortening the response window

An alert is only useful if someone can turn it into a decision before the attacker has moved on. As automation compresses the time between steps, that handoff becomes a security control in its own right.

What the September report described

In its September 8 report, Google Threat Intelligence Group described adversaries progressing from individual prompts to agentic workflows and AI-enabled automation. It reported a Q2 campaign in which attackers compromised a cloud resource, then planned and executed agent-enabled credential harvesting in under six hours. Read the GTIG report.

The report also covered attempts to manipulate AI coding assistants and security scanners during software supply-chain attacks, alongside theft of AI-related credentials and misuse of cloud resources. These were observed examples, not a claim that every attack now runs autonomously.

Find the waiting time in your response

My recommendation is to trace one realistic incident through your team. A provider detects an unusual login. Who receives the alert? Who decides whether to disable the account? Who has the access to do it? What happens outside normal business hours?

Time those handoffs in a short exercise. The useful measurement is how long it takes to reach an authorized person and complete the containment step. A large collection of alerts will not help if each urgent decision waits in an unattended mailbox.

Prepare a few actions in advance

Choose common situations and agree the first response with the business owner. Examples include a suspicious privileged session, an unexpected remote-access tool, or unusual use of a cloud API credential. Write down the evidence required, who can act, and how to confirm the action worked.

Keep these procedures specific to your systems. Review the operational impact before automating a response that could disconnect staff or interrupt a customer service. Rehearse the recovery path as well as the containment action.

Reduce what a stolen identity can do

Review privileged and service accounts, unused permissions, and credentials that never expire. Separate identities used for development, automation, and administration. An attacker who obtains one credential should encounter another enforceable limit before reaching an unrelated system.

For an AI workflow, include the model service, its tool connections, and the data it can retrieve. Check usage and spending limits on cloud accounts too. Unauthorized workloads can create a business problem even when the attacker does not steal a customer database.

Keep review in the development path

A coding assistant's suggestion still needs dependency review, testing, and a person responsible for release. Make the package source and the access requested visible during review. Treat instructions found in a repository or document as untrusted input to the workflow.

The practical response to faster attacks is a response path with fewer uncertain handoffs, smaller permissions, and evidence that containment works. Start with one scenario your team can rehearse this month, then use the result to improve the next one.