When even Google admits it doesn’t have AI security figured out, we should all pay attention. A TechCrunch report from this weekend paints a sobering picture of the current security landscape around AI systems.
23 minutes of ghost time
Security firm Aikido found a troubling gap: when developers delete a compromised Google API key, it stays active for up to 23 minutes. The reason: Google’s revocation mechanism propagates gradually across its infrastructure. During that window, attackers can exfiltrate files and cached conversation data from Gemini.
Sounds like a minor issue? It’s not. In the world of modern attacks, the average time between initial breach and handoff to the next stage has dropped to 22 seconds. Twenty-three minutes is an eternity.
Five-figure bills overnight
Even more unpleasant: Google Cloud developers have been surprised with five-figure bills caused by unauthorized API calls to Gemini. If someone grabs your API key and uses it for massive AI requests, you’re the one paying.
Shadow agents - the invisible risk
A new phenomenon is making security teams particularly nervous: shadow agents. These are AI agents that employees deploy independently to automate tasks. The problem: these agents create invisible, uncontrolled data pipelines and can expose organizations to security, compliance, and privacy risks - without IT even knowing about it.
What LinkedIn’s CISO says
Lea Kissner, Chief Information Security Officer at LinkedIn, puts it bluntly: the industry won’t understand AI security in a sustainable long-term way for ‘at least several years.’ That’s an honest assessment - and one worth taking seriously.
Google itself has responded: three new Security Operations agents are entering preview, building on an existing triage agent that processed over 5 million alerts in the past year. Analysis time per alert dropped from 30 minutes to 60 seconds. Progress, yes - but against 22-second attacks, even that’s not enough.
What this means for Claude users
Anthropic has taken a different approach with Project Glasswing and the Mythos model: using AI for defense rather than just reacting. But the fundamental problems - API key management, shadow agents, the speed of modern attacks - affect the entire industry. If you’re using AI tools in your organization, treat your API keys like passwords and keep an eye on uncontrolled AI agents.
Sources: