AI Sec Watch

At Anthropic's developer event, nearly half the attendees reported shipping code written entirely by Claude (an AI assistant), with many not reading it first before deploying it live. The article discusses how AI coding tools are becoming increasingly capable and how developers are automating their work, though not everyone agrees this approach is beneficial.

SpaceX, OpenAI, and Anthropic are planning major initial public offerings (IPOs, where companies sell shares to the public for the first time) in 2026, with SpaceX targeting a $1.75 trillion valuation. However, analysts warn these mega-cap floats resemble the late-1990s dot-com bubble, noting that all three companies are unprofitable and have opaque business models, with SpaceX's only profitable division being its Starlink internet service while its AI and space divisions operate at significant losses.

AI security strategies often fail in operational technology (OT) environments, like power plants and factories, because critical legacy systems don't send data to AI systems—a maintenance laptop running unpatched Windows 7 is common. AI trained on typical IT data (like web traffic logs) often misclassifies normal industrial traffic as threats, and automated responses can accidentally shut down production lines faster than actual attacks, because in OT systems availability (keeping things running) is more important than the IT security priorities of confidentiality and integrity.

Google is integrating CodeMender, an AI agent that automatically finds and fixes software vulnerabilities, into its larger Agent Platform ecosystem rather than keeping it as a standalone tool. CodeMender uses Gemini reasoning models (advanced AI that can think through complex problems) to analyze code vulnerabilities, generate fixes, and test them before showing them to developers. This shift suggests Google believes enterprises want autonomous security tools embedded within a governed infrastructure framework with identity and monitoring systems, rather than as isolated products.

OpenAI's Codex, an AI tool that helps developers write and manage code, has been recognized as a Leader by Gartner in enterprise coding agents. Codex goes beyond simple autocomplete (where an AI completes code as you type) by letting developers delegate complex tasks like understanding large codebases, running tests, and preparing work for human review while maintaining security and governance controls. The recognition highlights Codex's strengths in enterprise features like approval gates, RBAC (role-based access control, which limits what different users can do), sandboxing (isolating code in a safe environment), and audit trails.

Network-AI v5.4.4 has a critical authentication bypass where the MCP server (a tool that lets AI models call external functions) defaults to an empty secret, causing all authentication checks to pass unconditionally. Additionally, the server allows requests from any origin (CORS wildcard), so an attacker can trick a user into visiting a malicious website that sends commands to their localhost Network-AI server without needing any password, potentially invoking dangerous tools like config_set and agent_spawn.

Boxlite, a sandbox service for running containers, has a path traversal vulnerability (a security flaw where attackers can access files outside intended boundaries) in how it extracts container images. When processing tar files (compressed archives), Boxlite doesn't validate symlink targets (shortcuts to files or directories), allowing an attacker to create a malicious container image that writes files anywhere on the host system, potentially leading to remote code execution (running unauthorized commands on the computer).

BoxLite is a sandbox service that runs untrusted code in lightweight virtual machines (VMs, which are isolated computing environments). It claims to protect host files by mounting directories in read-only mode (preventing writes), but the vulnerability bypasses this: BoxLite tells the underlying VM system (libkrun) to mount directories without actually enforcing read-only restrictions, and it doesn't limit container capabilities (special permissions), so malicious code can remount directories as read-write and modify files that should be protected.

Pydantic AI had a security flaw where attackers could bypass protections against accessing cloud-metadata endpoints (special internal servers that store sensitive credentials) by encoding the IP address in IPv6 transition forms (IPv4-mapped IPv6, 6to4, or NAT64, which are ways to represent IPv4 addresses using IPv6 format). This flaw only affects applications that explicitly allow local file downloads with the `force_download='allow-local'` setting on URLs that could be influenced by untrusted users.