AI Sec Watch

A Chinese cybersecurity company called 360 Digital Security Group claims to have discovered 1,000 vulnerabilities (weaknesses in software that attackers can exploit) using AI tools, including some vulnerabilities found at the Tianfu Cup hacking contest. The article compares these claims to myths about Claude (an AI system), suggesting skepticism about the actual capabilities being reported.

Google announced new AI agents and security tools designed to help security teams keep pace with the increasing number of vulnerabilities and cyber threats. The company introduced three new agents embedded in Google Security Operations (for threat hunting, detection engineering, and gathering external intelligence), expanded the Wiz security platform to monitor AI development across multiple clouds, and created tools like AI-BOM (AI bill of materials, an inventory of all AI components used in an organization) and Agent Gateway to secure interactions between AI agents. These moves represent a shift toward automated, agent-based defense rather than relying solely on human analysts.

Google announced new AI agents and security tools designed to help security teams defend against AI-based attacks, particularly in response to threats like Anthropic Mythos. The company introduced three new agents within Google Security Operations to automate threat detection and response, expanded the Wiz platform to provide visibility across multiple cloud environments and AI development tools, and created new security measures like AI-BOM (a system that catalogs all AI components used in an organization) and Agent Gateway to govern how AI agents interact with each other and enforce security policies.

Trailmark is an open-source library that converts source code into a queryable call graph (a visual map of how functions and classes connect to each other) that AI systems like Claude can analyze directly. Rather than examining code as flat lists of findings, Trailmark lets AI reason about code structure as a graph, making it better at identifying security risks like whether untrusted input can reach vulnerable code.

Microsoft is releasing Agent Mode (previously called 'vibe working') in Office applications like Word, Excel, and PowerPoint, which is a more advanced version of Copilot (an AI assistant) that can actively perform tasks in documents rather than just answer questions. Previously, the AI models weren't powerful enough to let Copilot directly control applications, so it could only provide passive help like answering user questions.

Anthropic's Project Glasswing uses an AI model called Mythos that is extraordinarily effective at finding software vulnerabilities, discovering bugs that humans missed for decades and even chaining multiple bugs together into working exploits. However, the critical problem is that fewer than 1% of vulnerabilities Mythos finds are actually patched, revealing a massive gap between how fast AI can discover security flaws (machine speed) and how fast human teams can fix them (calendar speed, typically four days per cycle).

GPT-5.5 is a new AI model from OpenAI designed to handle complex work tasks like coding, research, and document creation with less user guidance than previous models. OpenAI conducted extensive safety testing including red-teaming (simulated attacks by security experts to find vulnerabilities) and feedback from nearly 200 early partners before release, and deployed it with what they describe as their strongest safeguards to date.

OpenAI released GPT-5.5, a more intelligent AI model that can handle complex, multi-step tasks like coding, research, and data analysis with less human guidance than previous versions. The model matches the speed of its predecessor while performing at a higher level and using fewer tokens (individual pieces of text that the AI processes). OpenAI says it tested GPT-5.5 with safety experts and external reviewers before release to reduce misuse risks.

Researchers at Palo Alto Networks built an autonomous multi-agent AI system called Zealot to test whether AI could independently perform cloud attacks. The system successfully chained together multiple exploitation techniques (SSRF, credential theft, and data theft) against a test Google Cloud environment, demonstrating that AI acts as a force multiplier for known cloud misconfigurations rather than creating entirely new vulnerabilities.