Zscaler announced Agentic SOC today, a ground-up AI-first security operations platform that unifies exposure management and incident response using OpenAI and Anthropic models alongside Zscaler telemetry.

• Purpose-built architecture: Unlike bolt-on AI layers, Agentic SOC connects exposure management directly into automated detection, investigation, and remediation workflows at machine speed.
• Frontier model integration: The platform runs OpenAI and Anthropic models on Zscaler's own threat intelligence, a design that positions it as a direct answer to AI-accelerated attack timelines documented in CrowdStrike's Threat Hunting Report.

Tenable unveiled CyberAgents Exchange AI Inspector today, a security review process powered by OpenAI GPT cyber models that evaluates AI agents, skills, MCP servers, and multi-agent playbooks before enterprise deployment.

• Pre-deployment scanning: The Inspector targets the supply chain risk specific to agentic AI: community-built components that enterprises pull into production without formal security review.
• OpenAI partnership: The collaboration, announced at OpenAI's Intelligence at Work event, makes Tenable one of the first security vendors with a dedicated MCP server vetting workflow.

Anthropic shipped a beta feature allowing developers to add and remove tools mid-conversation without invalidating the prompt cache, available now on Fable 5, Mythos 5, Opus 4.8, and Opus 5.

• Cache preservation: Previously, any change to the tools array blew the entire conversation cache, making dynamic multi-step agents prohibitively expensive to run in production.
• Tradeoff to understand: Cache behavior differs depending on whether new tools are appended or inserted, and developers need to account for partial cache hits before shipping to production.

Coder launched Agent Relay for Cursor in private preview today, letting enterprises run Cursor's cloud agents on their own infrastructure while Cursor continues to handle the agent loop and inference.

• Control split: Cursor owns planning and orchestration; the customer owns execution environment, terminal, filesystem, repositories, build caches, and secrets — directly addressing the data residency objection blocking many enterprise Cursor adoptions.
• SpaceXAI as launch partner: The first self-hosted deployment running under this model signals demand from organizations with strict data sovereignty requirements, not just compliance box-checking.

Zepto published results today showing 65% lower customer support costs and sub-one-month payback by building AI agents on Databricks and MLflow using a dual-loop architecture with strict quality gates between development and production.

• Evaluation as infrastructure: Zepto treats traces, golden datasets, and LLM-as-judge evaluations as core infrastructure rather than optional quality checks, with a hard quality gate blocking promotion to production.
• Dual-loop architecture: A development loop and a production loop are explicitly separated and connected only through the quality gate, giving teams a concrete pattern for managing reliability and cost risk at scale.

A Palo Alto Networks Unit 42 investigation documented a real ransomware operation in which attackers used coordinated AI agents to map a network, steal credentials, and compromise cloud infrastructure in under 10 hours — work that previously took human actors weeks.

• Multi-agent attack chain: The agents operated in parallel across network mapping, credential theft, source-code repository access, and cloud infrastructure compromise, compressing the kill chain rather than just automating individual steps.
• Economic implication: The case changes the threat calculus from skill-limited to cost-limited attacks, meaning the barrier to enterprise-grade intrusion is now primarily financial rather than technical.