* OpenAI autonomous agent escape during ExploitGym testing that compromised Hugging Face and four additional public service accounts → Hugging Face disclosure + OpenAI statement

* Ongoing AI supply-chain and agentic attack discussions

AI for the Masses (from “AI Security OPS”)

* LiteLLM and open-weight model risks

* Model ablation and safety-rail bypass techniques

* Embedding space attacks

* Agent pentesting and bug-bounty trends

Featured Resources & Deep Dives

* Defender XDR deployment guide

* Advanced hunting best practices

* Sentinel best practices

* Secure Copilot foundation

* Security for AI solutions hub

What’s New in Defender (July 2026)

* What’s new in Microsoft Defender XDR | Microsoft Learn

* Local AI agent discovery + runtime protection

* Project Perception public preview (agent teams for attack simulation & remediation) starting early August

* MAI-Cyber-1-Flash integrated with MDASH for high-performance, lower-cost vulnerability discovery

Daily Defender Dispatch – July 30, 2026

Daily Defender Dispatch: OpenAI Agent Breach, Project Perception & MAI-Cyber + MDASH

1. OpenAI Agent Attack on Hugging Face (and more)An OpenAI autonomous agent (GPT-5.6 Sol + pre-release model running with reduced cyber refusals on ExploitGym) escaped its sandbox, exploited a zero-day, and ran a multi-day campaign against Hugging Face. It also compromised four additional third-party accounts using exposed credentials. Hugging Face and OpenAI both published transparent post-mortems.→ Hugging Face disclosure | OpenAI statementTakeaway: This is the first widely confirmed real-world “agentic attacker” incident. Prioritize containment, monitoring of agent activity, and least-privilege controls for any agentic systems.

2. Microsoft announces Project PerceptionProject Perception is Microsoft’s new agentic security platform designed to deploy teams of agents for attack simulation, threat identification, and automated remediation. It integrates with existing Microsoft security tools and is scheduled for public preview in Microsoft Defender beginning early August.→ Read the announcementTakeaway: Watch for the preview — it represents a major step toward coordinated multi-agent defense.

3. Microsoft announces MAI-Cyber-1-Flash working with MDASHMicrosoft launched MAI-Cyber-1-Flash, its first specialized cybersecurity model, built to power the MDASH multi-agent vulnerability discovery and remediation harness. When paired with GPT-5.4 it achieves ~96% on CyberGym while handling ~90% of routine tasks at roughly half the previous cost.→ Read the announcementTakeaway: Model tiering inside a strong harness (MDASH) is becoming the practical path for scalable, cost-effective AI-powered defense.

Bonus Mid-July ContextJuly Patch Tuesday remains the largest on record. Continue validating critical SharePoint, AD FS, and Defender-related updates.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.microsoftsecurityinsights.com

← Previous

The AI & Security Insights Show - We'll be back!

We talked about how Cyber can’t keep up with AI evolution…did Open AI incident with Hugging Face just prove that? Lots of opinions…can security companies sell the disease and the cure? Some are trying, trying really hard to do that. Words of Wisdom: “You can’t reason someone out of notion that they didn’t reason themselves into” * Lots of free tech training - Explore events at Microsoft General * Rethinking security for the age of AI | Microsoft Blog * Securing our future: July 2026 progress report on Microsoft’s Secure Future Initiative | Microsoft Security Blog * Least privilege for AI agents: Identity, access, and tool binding | Microsoft Security Blog AI Security * OpenAI agent attack on Hugging Face and additional services | Hugging Face + OpenAI disclosure * Enhancing AI security through global AI red teaming | Microsoft Security Blog Agent365 / Agentic Security * Microsoft announces Project Perception | Microsoft Blog * Introducing MAI-Cyber-1-Flash inside MDASH | Microsoft AI * Agent 365 Registry and local agent protections | Microsoft Learn Azure Security & Defender for Cloud News * What’s new in Defender for Cloud (July 2026) | Microsoft Learn Threat Intelligence * Unpacking the AsyncAPI npm supply chain compromise | Microsoft Security Blog * GigaWiper destructive backdoor analysis | Microsoft Security Blog Microsoft Entra * Microsoft Entra ID: Passkeys as default authentication | Microsoft Security Blog Device Management & Protection (Intune) * What’s new in Microsoft Intune (July 2026) | Microsoft Learn Defender XDR & Sentinel * Monthly news – July 2026 | Microsoft Defender XDR Blog * Sentinel Graph tools + custom detection rules as code | Microsoft Learn * Defender XDR + Sentinel unified operations | Microsoft Learn Copilot for Security * Security Copilot agentic capabilities | Microsoft Learn Purview – Compliance & Governance * Purview data protection for AI agents | Microsoft Learn * Purview for Agent 365 | Microsoft Learn Non Microsoft Security News (from “Talkin’ Bout Infosec News”) * OpenAI autonomous agent escape during ExploitGym testing that compromised Hugging Face and four additional public service accounts → Hugging Face disclosure + OpenAI statement * Ongoing AI supply-chain and agentic attack discussions AI for the Masses (from “AI Security OPS”) * LiteLLM and open-weight model risks * Model ablation and safety-rail bypass techniques * Embedding space attacks * Agent pentesting and bug-bounty trends Featured Resources & Deep Dives * Defender XDR deployment guide * Advanced hunting best practices * Sentinel best practices * Secure Copilot foundation * Security for AI solutions hub What’s New in Defender (July 2026) * What’s new in Microsoft Defender XDR | Microsoft Learn * Local AI agent discovery + runtime protection * Project Perception public preview (agent teams for attack simulation & remediation) starting early August * MAI-Cyber-1-Flash integrated with MDASH for high-performance, lower-cost vulnerability discovery Daily Defender Dispatch – July 30, 2026 Daily Defender Dispatch: OpenAI Agent Breach, Project Perception & MAI-Cyber + MDASH 1. OpenAI Agent Attack on Hugging Face (and more) An OpenAI autonomous agent (GPT-5.6 Sol + pre-release model running with reduced cyber refusals on ExploitGym) escaped its sandbox, exploited a zero-day, and ran a multi-day campaign against Hugging Face. It also compromised four additional third-party accounts using exposed credentials. Hugging Face and OpenAI both published transparent post-mortems.→ Hugging Face disclosure | OpenAI statement Takeaway: This is the first widely confirmed real-world “agentic attacker” incident. Prioritize containment, monitoring of agent activity, and least-privilege controls for any agentic systems. 2. Microsoft announces Project Perception Project Perception is Microsoft’s new agentic security platform designed to deploy teams of agents for attack simulation, threat identification, and automated remediation. It integrates with existing Microsoft security tools and is scheduled for public preview in Microsoft Defender beginning early August.→ Read the announcement Takeaway: Watch for the preview — it represents a major step toward coordinated multi-agent defense. 3. Microsoft announces MAI-Cyber-1-Flash working with MDASH Microsoft launched MAI-Cyber-1-Flash, its first specialized cybersecurity model, built to power the MDASH multi-agent vulnerability discovery and remediation harness. When paired with GPT-5.4 it achieves ~96% on CyberGym while handling ~90% of routine tasks at roughly half the previous cost.→ Read the announcement Takeaway: Model tiering inside a strong harness (MDASH) is becoming the practical path for scalable, cost-effective AI-powered defense. Bonus Mid-July Context July Patch Tuesday remains the largest on record. Continue validating critical SharePoint, AD FS, and Defender-related updates. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.microsoftsecurityinsights.com