AI Sec Watch

Amazon is launching Alexa for Shopping, an LLM-powered AI assistant (a language model trained to understand and respond to text) that integrates directly into Amazon.com to answer shopping questions and provide product recommendations. This new assistant replaces Amazon's previous shopping AI called Rufus and will appear prominently in the Amazon app and website when users type queries.

Microsoft developed MDASH, an AI system that uses over 100 specialized AI agents to automatically find software vulnerabilities, and it discovered 16 previously unknown Windows flaws, including four critical RCEs (remote code execution attacks where attackers can run commands on a system). The vulnerabilities were patched in Microsoft's May 12 Patch Tuesday release, and the system will enter private preview for enterprise customers next month.

Palo Alto Networks launched Idira, a new identity security platform designed to manage and secure human users, machine identities (non-human accounts that systems use), and AI agents as enterprises increasingly deploy autonomous AI systems. Unlike traditional identity management systems, Idira treats all identities as privileged and uses dynamic privilege controls, continuously discovering identities across cloud and developer environments while dynamically granting and revoking access permissions in real time rather than using static access tokens.

The UK's AI Security Institute tested whether GPT-5.5 (OpenAI's model) could find security vulnerabilities as well as Claude Mythos, and found they perform comparably. A smaller, cheaper model also performed equally well, though it needs more scaffolding (additional structure and guidance provided by the person writing prompts to the AI).

Codex, OpenAI's coding agent, previously lacked a sandbox (a constrained execution environment with restricted permissions) on Windows, forcing users to either approve every command or allow unrestricted access. To solve this, the Codex team built a custom sandbox implementation because existing Windows tools like AppContainer, Windows Sandbox, and Mandatory Integrity Control labeling were either too restrictive for open-ended developer workflows or incompatible with the product requirements.

CISA and G7 cyber agencies released guidance on minimum elements for AI software bills of materials (SBOMs, documents listing all components and dependencies in software), helping security leaders assess AI system risks before deployment. Unlike traditional SBOMs that only track code, AI SBOMs must document models, training data, prompts, infrastructure, and other AI-specific elements because AI systems' behavior depends on data and models as much as code. The guidance gives organizations a framework to ask vendors for transparency during procurement, though it shows what vendors claim exists rather than proving the systems are trustworthy.

A legal dispute between Elon Musk and OpenAI leaders has revealed that private conversations shared with AI chatbots are not actually confidential, as evidenced by Greg Brockman's personal messages about Musk being used as court evidence. This case demonstrates that users should be cautious about disclosing sensitive personal information to AI systems because that data may not remain private.

SoftBank's Vision Fund recorded a $46 billion gain in the past year, mostly from its massive investment in OpenAI, which has grown to be worth $852 billion. However, this heavy concentration of OpenAI in SoftBank's portfolio has raised concerns about the company's debt and financial stability, leading ratings agency S&P Global to downgrade its outlook from stable to negative.

OpenAI discovered that two employee devices were compromised by malware hidden in a TanStack npm package (a JavaScript library downloaded from an online repository) as part of a broader supply chain attack called Mini Shai-Hulud. The attackers gained limited access to internal source code repositories and exfiltrated some credentials, but OpenAI found no evidence that customer data, production systems, or intellectual property were compromised. OpenAI responded by isolating affected systems, revoking credentials, rotating code-signing certificates (the digital signatures that verify software is authentic), and working with platform providers to prevent misuse of the compromised certificates.