This Week in AI Security - 5th March 2026

In this week's episode, Jeremy records straight from the sidelines of the [un]prompted security conference in San Francisco. Before diving into his key takeaways from the event, he covers a massive, AI-assisted data breach and a critical shift in how Google API keys must be handled.

Key Stories & Developments:

• Nation-State AI Hack: A hacker reportedly used Anthropic’s Claude to identify vulnerabilities and OpenAI’s GPT-4.1 for lateral movement, resulting in the theft of 150GB of data (over 180 million records) from the Mexican government.
• MCP Infrastructure Flaws: An unauthenticated Server-Side Request Forgery (SSRF) flaw leading to Remote Code Execution (RCE) was found in a widely used Atlassian MCP.
• The Gemini API Key Crisis: A flaw in the Gemini AI panel allowed browser extensions to escalate privileges. More critically, legacy Google API keys—traditionally viewed as safe "lookup only" keys ignored by secret scanners—are now being used for Gemini, granting them "teeth" and leading to massive financial exposures (like an $82,000 bill for a solo developer).

Dispatches from the Unprompted Conference: Jeremy shares his top thematic observations from the event, including:

• The "Zero-Day Clock": The mean time to exploit availability has plummeted from months to mere hours. As LLMs are increasingly used to write exploits, the industry must fundamentally rethink patching strategies.
• LLMs Finding Legacy Bugs: Researchers demonstrated LLMs uncovering vulnerabilities in massive software projects that have evaded human detection for decades—some predating the invention of Git.
• Treating Prompts as Code: A key takeaway from Google's Gemini workspace team: as prompts become the primary instruction set for executing tasks, developers must apply traditional secure coding hygiene and logic to their prompt engineering.

Episode Links

https://www.bloomberg.com/news/articles/2026-02-25/hacker-used-anthropic-s-claude-to-steal-sensitive-mexican-data

https://blog.pluto.security/p/mcpwnfluence-cve-2026-27825-critical

https://cyberpress.org/critical-servicenow-ai-platform-flaw-allows-remote-code-execution-attacks/

https://www.darkreading.com/endpoint-security/bug-google-gemini-ai-panel-hijacking

https://trufflesecurity.com/blog/google-api-keys-werent-secrets-but-then-gemini-changed-the-rules

https://boingboing.net/2026/02/27/stolen-gemini-api-key-racks-up-82000-in-48-hours-for-solo-dev.htmlhttps://unpromptedcon.org/

Worried about AI security? Get Complete AI Visibility in 15 Minutes. Discover all of your shadow AI now. Book a demo of Firetail's AI Security & Governance Platform: https://www.firetail.ai/request-a-demo