AI Sec Watch

Security researchers found a prompt injection flaw (tricking an AI by hiding instructions in its input) in Google's Antigravity IDE that could bypass its Secure Mode sandbox protections and achieve RCE (remote code execution, where an attacker can run commands on a system they don't own). The vulnerability came from insufficient input validation in the file search tool's Pattern parameter, allowing attackers to inject malicious command-line flags that converted a simple file search into arbitrary code execution. Google acknowledged the issue in January and fixed it internally, and Antigravity users are now protected without needing to take action.

ChatGPT Images 2.0 is an updated image generation model (software that creates pictures from text descriptions) with better ability to render text within images, support for multiple languages, and improved visual reasoning (the ability to understand and analyze images). The announcement introduces new features but does not discuss security issues or problems requiring fixes.

Google patched a vulnerability in Antigravity, its agentic integrated development environment (IDE, a coding tool that can take autonomous actions), that allowed attackers to execute arbitrary code through prompt injection (tricking an AI by hiding instructions in its input). The flaw combined the tool's file-creation abilities with insufficient input validation in its find_by_name search function, letting attackers inject malicious commands that bypassed Antigravity's Strict Mode security restrictions.

AI company Anthropic announced it created a powerful model called Mythos Preview that can find and exploit software vulnerabilities (weaknesses that attackers could use), and decided not to release it publicly due to concerns about risks to economy, safety, and national security. However, some experts question whether the model is actually as capable as Anthropic claims, and the decision raises questions about whether this move is genuine responsibility or a publicity strategy.

Organizations typically have far more AI tools running than they realize, including unapproved ones that bypass traditional security controls, a problem called shadow AI (unauthorized AI usage that goes undetected). CrowdStrike's new Shadow AI Visibility Service addresses this by using telemetry-based evidence (data collected from system monitoring) to discover both approved and unapproved AI across endpoints, cloud, and SaaS environments, since most security teams lack visibility into their actual AI footprint.

Claude Code, an agentic coding tool (AI that can write and execute code), had a sandbox escape vulnerability before version 2.1.64 where sandboxed processes could create symlinks (shortcuts pointing to files outside their designated area) that allowed writing to locations outside the workspace without user permission. An attacker could exploit this by injecting malicious instructions into Claude Code's input, potentially executing code outside the intended sandbox.

LlamaIndex v0.14.21 is a maintenance release that fixes several bugs in the core library, including a KeyError (an error when looking up a value in a data structure that doesn't exist) in the DocumentSummaryIndex deletion function, handling of output formatting errors, and UTF-8 encoding issues in file operations. The release also updates dependencies across many embedding and indexing modules to keep the library's supporting code current.

Codex, an AI tool that generates code and assists with software development tasks, has grown from 3 million to 4 million weekly users and is now being adopted by major enterprises like Virgin Atlantic, Notion, and Cisco to speed up development workflows. OpenAI is expanding Codex adoption through a program called Codex Labs, which provides expert guidance to organizations, and by partnering with global consulting firms (like Accenture and Infosys) to help enterprises integrate Codex into their software development processes at scale.

Amazon is investing up to $25 billion more in Anthropic, an AI company known for its Claude AI models (large language models, or LLMs, which are AI systems trained on vast amounts of text to generate human-like responses), on top of an earlier $8 billion investment. As part of this deal, Anthropic will spend over $100 billion on Amazon's cloud services and custom AI chips over the next decade to expand its computing capacity (the processing power needed to train and run AI models). Anthropic made this agreement because its infrastructure has been strained by rapidly growing demand from enterprise customers and users of Claude.