In this week's episode, Jeremy covers seven stories that highlight the continuing pattern of API-level risks, the rise of multi-agent threats, and new academic insights into LLM fundamentals.

Key stories include:

• RCE via PyTorch: A high-severity vulnerability (with an assigned CVE) was discovered in the widely-used PyTorch package, enabling Remote Code Execution (RCE) through malicious payloads at the API layer. This reinforces the trend of the API being the primary attack surface for AI applications.
• AI Browser Local Command Execution: Researchers found an API flaw in AI browsers that allowed a malicious instruction set to execute local commands on a user's machine via an embedded extension.
• Klein Bot Vulnerabilities: An open-source coding agent was found to have multiple security flaws, including the exfiltration of API keys and the disclosure of its underlying model (Grok), validating OWASp's risk categories.
• Multi-Agent Risk in ServiceNow: Researchers demonstrated that in ServiceNow’s new A-to-A agentic workflows, default configurations place agents in the same network, allowing them to communicate and be exploited using the privileges of the human user who created them.
• The "Subspace Problem" of Red Teaming: Academic research argues that current LLM red teaming methods are flawed because they test human language, not the numerical token strings the LLM actually processes, meaning predictable token-level vulnerabilities remain hidden.
• AI Evaluation Shift: A paper argues that non-deterministic LLM environments require a shift away from binary "yes/no" security checks (like traditional network security) toward scenario-based testing for better risk evaluation.
• Positive ROI of AI in Security: A Google paper provides positive data for early movers, showing that AI can triage at least 50% of security incidents, leading to reduced human workloads and faster response times, providing a strong case for simple, prompt-based AI improvements in security operations.

------

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