- SMB/SME
- Enterprise
· ASC Editorial Team
Stolen AI Keys Are Stolen Cloud Keys: What Google's Latest Threat Report Means for Your Controls
On 16 November 2026, Google Threat Intelligence Group and Mandiant published their latest report on how attackers use AI. It describes the first case they have seen of attackers using AI to find and exploit vulnerabilities with no human involved. On an exposed server, they found a system that compares each new web application patch with the old code, drafts an exploit for the flaw it fixes, and runs it. They also saw several groups using open-source AI penetration-testing tools to automate parts of real intrusions.
Google adds two caveats. These tools currently trade stealth for speed, so they leave traces defenders can spot. And many exploits written by AI don't work. The direction is still clear. Google counted 11 zero-day exploits a month in 2026, up from 8 a month in 2025, with 22 in August alone. We wrote about how quickly attackers move once inside in our ARTEX article. The time between a patch and an attack is shrinking as well. That is one more reason to limit what attackers can reach at all, as we argued in Patched but Exposed.
For most organisations, the more immediate risk is in a different part of the report: the theft of AI accounts and keys, which the industry calls LLMjacking. Attackers steal access to an organisation's AI services and either use it themselves or resell it. In Mandiant's investigations this year, charges started climbing within hours. Some fraudulent charges passed $100,000 in a single day, and in the worst cases the total reached $1 million before anyone stopped it.
The report describes how the access is stolen: - keys left in public code repositories; - session tokens taken through fake websites, or from staff computers and build pipelines; - credentials bought from criminals who run password-stealing malware.
Once in, attackers created new service accounts or gave existing ones administrator rights, so that they kept their access after a key was revoked. In one case they used a flaw in LiteLLM, a popular gateway for AI services (CVE-2026-40217), to turn a limited key into full administrator access.
An AI key is a cloud credential with a cost meter attached. It often has broad rights, and it is frequently created by a developer or a project team, outside the controls that cover other privileged accounts. Treat it as a privileged secret.
Four checks follow.
First, list every AI service your organisation uses: the main providers, cloud AI platforms, gateways, and trials someone started and forgot. Note who owns each key and what it can do.
Second, limit each key to the models and projects it needs, and set spending limits and alerts that someone will actually see the same day.
Third, keep keys out of code. Store them in a secrets manager, scan your repositories for leaked keys, and revoke keys when staff or suppliers leave.
Fourth, alert on the signs the report describes: new service accounts, administrator rights being added, requests for more computing quota, and sudden jumps in AI usage. If you run an AI gateway such as LiteLLM, patch it like any other internet-facing system.
If one of your AI keys were being used by someone else tonight, how soon would you know?
Source: Google Threat Intelligence Group: Agentic Threat Activity Crosses a Significant Threshold
Get Insights by email
A short round-up of new articles every Monday morning. Nothing else.