AI Sec Watch

Pipecat's development runner has a path traversal vulnerability (a flaw that lets attackers access files outside the intended directory) in its `/files` endpoint. An attacker can use URL-encoded slashes (`%2F` instead of `/`) to bypass Starlette's (the web framework) security checks and read any file accessible to the Pipecat process, such as SSH keys or system files, without needing credentials.

Google updated its spam policy to classify attempts to manipulate its AI search results as spam, including tactics like biased listicles or recommendation poisoning (injecting false information to trick an LLM into giving preferred answers). This rule applies to Google Search's AI features like AI Overview and AI Mode.

OpenAI announced a new preview feature that will let ChatGPT connect directly to users' bank accounts through Plaid, a platform that links banking apps to third-party services. This integration would give the chatbot access to detailed financial information, including credit card debt and account balances, to help answer users' finance questions.

This cybersecurity news roundup covers several significant incidents and developments, including a data breach at Nvidia's GeForce NOW service in Armenia that exposed user personal information, extended security update timelines for foreign-made routers and drones, and OpenAI's offer to give EU regulators access to a specialized version of GPT-5.5 for monitoring cyber security risks. The roundup also highlights an active malware campaign targeting developers with fake Claude Code installers, an Iran-linked group breaching South Korean electronics manufacturers, and Google's Android 17 release introducing AI-driven security features like verified financial calls and real-time threat detection.

Agentic AI tools (AI systems that can plan, make decisions, and take actions without constant human supervision) are becoming more common in organizations but introduce significant security risks beyond traditional AI systems. These risks include broader system access, unpredictable behavior, and difficulty explaining AI actions. The NCSC and international partners recommend organizations adopt agentic AI carefully by starting with low-risk tasks, deploying incrementally with tight controls, maintaining human oversight, and ensuring clear human accountability before connecting agents to real systems or data.

OpenAI disclosed that two employee devices were compromised through the Mini Shai-Hulud supply chain attack on TanStack (a software dependency library), resulting in limited credential theft from internal code repositories but no user data or production systems were affected. Because the compromised repositories contained signing certificates (digital credentials that verify software authenticity) for macOS apps, OpenAI revoked the old certificates and requires macOS users of ChatGPT Desktop, Codex App, Codex CLI, and Atlas to update to the latest versions before June 12, 2026, when the old certificates will be blocked by macOS protections.

OpenAI disclosed that two employee devices were infected during a supply chain attack on TanStack, a web development framework, which allowed attackers to steal credential material from internal source code repositories. The stolen credentials gave attackers access to code-signing certificates (digital keys used to verify that software is authentic) for OpenAI's applications on iOS, macOS, Windows, and Android. OpenAI confirmed that no customer data or intellectual property was compromised, but took steps to prevent further risk.

MLflow (an open-source platform for managing machine learning workflows) versions 3.9.0 and earlier have a security flaw where certain API endpoints don't require authentication even when the server is set up with authentication enabled. This happens because the authentication check only protects `/gateway/` routes, leaving other endpoints like the Job API and trace ingestion API unprotected, allowing attackers to submit jobs, view results, and inject fake data without logging in.

datasette-llm-limits is a plugin that works with Datasette (a tool for exploring databases) to set spending limits on how much money users can spend on LLM API calls. The plugin lets administrators configure daily or rolling limits per user or globally, for example restricting one user to $1.00 of LLM usage per 24-hour period.