AI Sec Watch

OpenAI discovered crashes in Rockset, a C++ service that helps ChatGPT search for relevant data, where programs would mysteriously stop after functions returned to invalid memory addresses. After investigating like epidemiologists and analyzing crash data patterns, the team discovered two unrelated bugs: silent hardware corruption on an Azure server where the CPU made math errors, and an 18-year-old race condition (a bug where two processes interfere with each other) in GNU libunwind, a widely-used open source library for tracking function calls.

Meta contractors posed as minors online and sent thousands of harmful prompts (requests designed to trigger unsafe responses) to competitor chatbots like ChatGPT, Gemini, and Character.AI to test how they handled questions about suicide, self-harm, sex, and drugs. The project, called Cannes and managed by contractor Covalen, involved over 45,000 prompts sent through dummy accounts, but the competing companies were not informed of this testing. Meta defended the effort as routine safety testing, though the documents reviewed do not explain how the collected data was actually used.

OpenAI is releasing a hardware device on July 15th designed to work with Codex, its AI-powered coding tool. The device, created in partnership with Work Louder (a company that makes mechanical keyboards and customizable macro pads), appears to be a square pad with buttons that will provide shortcuts for Codex features.

Research by Boston University professor Emma Wiles shows that calling AI agents (software tools programmed to work in loops toward specific goals) "coworkers" or "employees" actually makes humans worse at their jobs: people caught 18% fewer errors and were 44% more likely to pass questionable work to managers instead of correcting it themselves when AI was framed as an employee rather than a tool. This framing inverts responsibility, making humans feel less accountable for AI outputs and creating unrealistic expectations, especially as AI agents get embedded in critical fields like healthcare and government.

Ornith-1.0 is a new open-source AI model designed for coding tasks, built on top of existing models (Gemma 4 and Qwen 3.5) and available in multiple sizes ranging from 9 billion to 397 billion parameters. The model is MIT licensed and achieves top performance on coding benchmarks among similarly-sized open-source models, with users reporting it works well for tasks like analyzing code and handling multiple tool calls in agent systems (where the AI can use external tools to complete tasks).

A vulnerability in Devolutions PowerShell Universal 2026.2.0 allows authenticated users with AI Agent read access to obtain sensitive authentication tokens (reusable credentials that grant access to systems) because these tokens are sent in plaintext (unencrypted, readable text) in job API (an interface for running automated tasks) responses. This could allow attackers to gain higher-level access than they should have.

Lawmakers are proposing a new version of the Health and Location Data Protection Act that would ban companies, including AI chatbots like ChatGPT and Claude, from selling Americans' health and location data to data brokers (third parties who buy and sell personal information). The updated bill expands on an earlier 2022 version to specifically address the AI era by preventing not just data brokers but other companies from selling this sensitive information.

Claude Code, a tool that uses AI to help write software, had a security flaw in versions 2.1.38 through 2.1.163 where it could be tricked into creating special folders named '.git' and accessing files outside its restricted sandbox (a controlled environment that limits what software can access). An attacker could exploit this by creating a malicious code repository and convincing a user to run Claude Code on it, potentially allowing the attacker to run commands on the user's computer outside the sandbox's protections.

Enterprise organizations are increasingly investing in agentic AI (AI systems that can autonomously plan and execute multi-step tasks) to automate workflows and improve business outcomes, with technology teams showing high confidence in agents for structured tasks like data monitoring and report generation. However, agent effectiveness drops when complex tasks require business context (the specific goals, rules, and data relevant to an organization), which remains difficult to supply at the speed and quality needed. Human oversight is emphasized as essential to deploying agentic AI safely and reliably.