This Week in AI Security - 4th December 2025

In this week's episode, Jeremy dissects two critical security issues and shares key strategic takeaways from the recent Ascent Community Summit on Advancing AI Security.

Security Incidents & Research:

• OpenAI Third-Party Breach: We examine the security incident where OpenAI was affected by a third-party breach via the Mixpanel analytics platform. While customer PII was exposed, prompt and data content was not impacted. Jeremy notes that the API was the attack surface, reinforcing a recurring theme in AI-related incidents.
• Adversarial Poetry: We break down a fascinating academic paper demonstrating that embedding malicious prompts inside poetry is a successful technique for bypassing LLM guardrails. In some models, this "adversarial poetry" increased the Attack Success Rate (ASR) by over 60%, showing how context manipulation can trick frontier models.

Ascent Community Summit Takeaways: Jeremy shares high-level insights from the summit (co-hosted by Paladin and Georgia Tech), focusing on securing critical sectors (Defense, Infrastructure, Healthcare). Key themes include:

• Core Requirements for AI: The need for math expertise, dedicated compute infrastructure, massive data access, and specialized people.
• The New Perimeter: Discussion shifted from "identity as the perimeter" to data being the key asset and central focus for security controls.
• Supply Chain Risks: The societal impact of the AI boom, including increased strain on electricity, cooling, and bandwidth for data center infrastructure.
• Brakes on a Fast Car: The CISO's role is framed as enabling maximum speed while having the ability to act as the "brakes on a very fast moving car" (Dundee West, GSK), emphasizing rapid response over stagnation.

Episode Links

1. https://openai.com/index/mixpanel-incident/
2. https://arxiv.org/pdf/2511.15304
3. https://sites.gatech.edu/asccent/summit/

------

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