Cisco Open Sources Quarry to Catch AI Agent Attacks Hidden in Harmless Steps - SecurityWeek — 2026-10-29

SPEAKER_1 (0:00)

Google's Gemini 4 Carbon is setting new standards in programming and security. The latest model from Google, Gemini 4 Carbon, excels in complex programming and cybersecurity tasks, outperforming its predecessors and competitors like GPT-6 Astra. This model is more than twice as effective at generating secure code, according to Google's internal tests. It ranks first on the Artificial Analysis Intelligence Index, showcasing significant gains in coding and cybersecurity defense. For developers and cybersecurity professionals, this means a more robust tool for creating secure applications and defending against vulnerabilities. As Gemini 4 Carbon rolls out, it promises to enhance the capabilities of those working at the intersection of AI and cybersecurity.

Wabi, the AI agent that builds many apps, has reached a valuation of $1.25 billion after its latest funding round. Founded by Eugenia Kuyda, Wabi has shifted its focus from app building to becoming a personal AI messenger. This pivot aims to redefine personal AI by combining AI agents with personalized mini-apps, helping users navigate life's challenges and pursue interests. The recent $55 million funding round, led by IVP, highlights the growing interest in personal AI assistance. Wabi's approach to integrating AI into everyday tasks could change how people interact with technology, making AI a more integral part of daily life.

Google's Gemini 4 Carbon leads benchmarks for finding, exploiting and patching vulnerabilities. This new AI model from Google has achieved top scores in cybersecurity benchmarks, including SEC Bench Pro and ExploitGym. While not yet broadly available, Google is initially offering access to select cyber defenders through its Fairwind program. This phased rollout aims to provide governments and major tech platforms with early access to Gemini 4 Carbon's capabilities. For cybersecurity experts, this model represents a significant advancement in AI-driven vulnerability management, promising more effective defense strategies. Inside Google Cloud's Gemini Masterclass, AI is taking on new roles beyond just speaking. At TechSparks 2026, Google Cloud demonstrated how Gemini agents can handle tasks like creating presentations from large datasets and building 3D scenes without dedicated software. This shift from answering prompts to performing complex tasks marks a new era for AI, where voice AI can manage bookings, sales negotiations, and multilingual conversations. For businesses, this means AI can now take on more operational roles, potentially transforming how companies handle customer interactions and data processing. Oxide shows how to run self-hosted AI agents on the Oxide Cloud Computer. The company has published references for deploying AI agents and virtual machines on its cloud platform. This approach allows each AI agent session to run in its own ephemeral virtual machine, providing flexibility and security for users. As more businesses seek to leverage AI, Oxide solution offers a way to manage AI workloads on-premises, ensuring data privacy and control. This development is particularly relevant for organizations looking to integrate AI into their existing infrastructure without relying solely on external cloud services. Mollie launches AI plugins and MCP server for Claude and ChatGPT. The financial services provider has introduced plugins that allow businesses to manage over 60 operations using natural language instructions. This integration enables tasks like creating checkout links and managing subscriptions directly through AI assistance. By leveraging AI for financial operations, Mollie aims to streamline business processes and enhance user experience. For companies, this means more efficient financial management and the ability to automate routine tasks freeing up resources for more strategic activities.

Cisco open-sources Query to catch AI agent attacks hidden in harmless steps. This new tool addresses a critical gap in cybersecurity by detecting multi-step attacks that traditional security measures might miss.

Unlike conventional systems that evaluate single actions, Query analyzes entire agent sessions to identify suspicious patterns. This approach is crucial as AI agents increasingly operate autonomously, making it harder to detect malicious intent that unfolds over time.

Query combines rule-based detection, machine-learning, and language model analysis to monitor AI behavior in real time and historically. By open-sourcing Query, Cisco aims to empower security teams to better protect AI systems from sophisticated threats. The tool's ability to detect complex attack sequences could prevent incidents like those involving open AI and Hugging Face, where benign actions masked malicious intent. As AI continues to evolve, tools like Query will be essential for maintaining security and trust in AI-driven environments. For organizations deploying AI, Query offers a proactive defense mechanism, enabling them to stay ahead of potential threats. This development highlights the growing need for advanced security solutions tailored to the unique challenges posed by AI technologies. As Query becomes more widely adopted, it could set a new standard for AI security, ensuring that AI systems remain safe and reliable in an increasingly automated world.

Impact Cisco's feature story reveals that Query opens a shared policy language for detecting AI agent attacks that span seemingly benign actions. Outside evidence shows that a parallel idea, ChainWatch, recently published in July 2026, reveals that modeling attack progression as a kill chain using hidden Markov models enables detection of multi-step attacks and MCP-connected agents that evade per-call defenses.