AI Sec Watch

Go's built-in fuzzing tool (a technique that tests software by feeding it random or semi-random inputs to find bugs) was missing key features available in other languages like Rust and C++, so researchers built gosentry, a modified version of the Go toolchain that adds stronger bug detection, grammar-based fuzzing (testing structured inputs like parsers), and coverage reports while keeping the same familiar interface. Gosentry detects bugs that vanilla Go fuzzing misses, including integer overflows (when numbers get too large for their storage), data races (when multiple threads access the same data unsafely), and goroutine leaks (abandoned lightweight threads), without requiring developers to rewrite existing fuzzing tests.

Agentic AI (AI systems that can independently execute tasks and take actions) is already running in many organizations without security teams fully understanding it, creating a significant security gap. Security professionals cannot effectively protect technology they don't understand, and teams that lack fluency in agentic AI are being bypassed by business units moving forward without their input. The article identifies three main categories of agentic AI risk: general-purpose coding agents (like GitHub Copilot), vendor-built agents using MCP (Model Context Protocol, which allows agents to connect to external services), and custom agents built by non-technical users, each requiring different security considerations.

Over 170 packages in popular NPM and PyPI repositories (code libraries that developers use) were compromised by the hacking group TeamPCP in a coordinated attack, including packages from TanStack, UiPath, and Mistral AI. The malware (malicious software) stolen sensitive information like API keys (credentials for accessing services), developer tokens, and cryptocurrency wallets, then tried to spread by using stolen GitHub tokens to publish infected versions of other packages. The attackers used a novel technique called a supply chain attack (compromising the tools and processes used to build and distribute software) by exploiting three security weaknesses in GitHub Actions (automated workflows for building and releasing code) to bypass security checks and make malicious packages appear legitimate.

CISOs (Chief Information Security Officers, the top security leaders at companies) are taking on increasingly important roles as AI becomes central to business operations and security threats grow. Companies like Brown & Brown and PayPal are addressing AI risks by creating AI security frameworks and governance structures that require security reviews before any AI tool is deployed, ensuring AI is used safely and responsibly.

TeamPCP compromised multiple popular software packages (from companies like TanStack, Mistral AI, and Guardrails AI) by injecting malicious code that steals credentials for cloud services, cryptocurrency wallets, and development tools. The attack used a technique called SLSA provenance (a system that verifies software was built securely) to make the malicious packages look legitimate, and the malware persists by modifying development tools like VS Code so it runs every time the tool starts.

OpenAI launched Daybreak, a new tool that uses AI models to help organizations find and fix software vulnerabilities before attackers can exploit them. Daybreak combines OpenAI's AI capabilities with Codex Security to automate tasks like code review, threat modeling, and patch validation. However, the article notes that AI tools have created a problem where vulnerabilities are discovered faster than developers can fix them, leading to 'triage fatigue' (where maintainers get overwhelmed sorting through many vulnerability reports, some of which may be false alarms generated by AI).

AutoScout24 Group, a large European and Canadian online car marketplace, implemented AI tools like ChatGPT and Codex (an AI coding assistant) across its 2,000 employees to speed up software development and improve code quality. By embedding Codex directly into engineering workflows and creating an "AI Champions" network for knowledge sharing, the company reduced development timelines from 2-3 weeks to 2-3 days for some projects while maintaining reliability.

NVIDIA engineers and researchers use Codex, an AI coding tool built on GPT-5.5, to automate complex engineering tasks and machine learning research workflows. The tool can work autonomously for long sessions, finding bugs and writing code that earlier models couldn't, and has enabled teams to build production systems and run experiments much faster than before. Codex integrates with remote infrastructure (SSH, or secure shell protocol, which lets users securely connect to distant computers) and can even test its own code as it's being built.

DeepChat, an open-source AI platform combining models, tools, and agents, has a vulnerability in versions before v1.0.4-beta.1 that allows remote code execution (RCE, where an attacker can run commands on a system they don't own). An attacker can use a malicious link in Markdown or a compromised AI endpoint to bypass security checks and execute arbitrary commands by exploiting unprotected pop-up window handlers in the application.