Mark Russinovich on AI Safety and Security at BlueHat

This is the exactly the kind of clarity the AI security conversation needs. Mark demos made the risks tangible, but the real value was in showing that practical defenses already exist. Tools like FIDES and RefChecker signal a shift toward engineering AI systems with security as a first‑class requirement, not an afterthought. AI safety is #cybersecurity now! The organizations that internalize this early will be the ones ready for what’s coming next in this AI world. #MSFTAdvocate #Microsoft #Employee #AISecurity #AISafety #Segurança #DigitalTransformation #AITransformation

It’s becoming increasingly clear that as we move from simple prompts to autonomous agents, the boundary between AI Safety and AI Security is disappearing. I had a great time this morning discussing this shift and why the risks we’ve been tracking, such as jailbreaks, indirect prompt injections, and hallucinations, are no longer just academic edge cases. As we grant agents greater access to our personal data, business systems, and decision making loops, these vulnerabilities become direct threats to our security posture. Understanding these risks is the first step toward building the resilient, secure by design agentic systems that the next era of computing requires.

Just attended one of the best keynotes I've seen in a while. The AI security talk at #BlueHat 2026 did not disappoint. Amazing talk Mark Russinovich 👏👏👏 In the village, the AeGIS AI Medusa's Memory Heist security game was very clever. A hands on challenge that tests your knowledge of AI vulnerabilities and how to defend against them. The kind of experience that makes you realize pretty quickly where your gaps are. #BlueHat #AISecurity #MSRC #AeGIS Microsoft Security Response Center Security conferences are really just reunions in disguise. #BlueHat 2026 was exactly that. Good vibes, familiar faces, and a few new ones. Already looking forward to MSRC reunion in Vegas!!!

I have been saying this for quite a while. This boundary is absolutely collapsing. Autonomous agents are a new risk vector, but time tested first principles are the key to managing the risk.

It’s great to see open discussion by thought leaders like Mark on the need to pay careful attention to the degree to which AI security breaches can lead to safety breaches.

Anthropic Increases API Usage Limits for Claude Models 📌 Anthropic is boosting API usage limits for its Claude models, removing previous bottlenecks for developers and enterprises. This expansion enables the seamless scaling of agentic workflows, allowing organizations to deploy high-throughput tasks like automated malware analysis and complex binary deobfuscation. By facilitating more intensive, iterative prompting, this update moves AI integration from experimental testing into robust, large-scale production environments. 🔗 Read more: https://lnkd.in/dtc3cgKP #Anthropic #Claudemodels #Apiusagelimits #Largelanguagemodels

Is your 5% CTR a performance win or a crime scene?🕵️♂️ In the programmatic ecosystem, high engagement without conversions is the ultimate red flag for SIVT. If you aren’t auditing the silent signs, you’re likely optimizing for botnets instead of humans. 👉Check how much ad fraud is costing you: https://lnkd.in/gUCQdvUS The Red Flags You Can’t Ignore: • The 3 AM Surge • Predictable Timing • Signature Loops Stop the Bleed Before the Bid. ClearTrust automates your IVT filtering with pre-bid protection to ensure your budget reaches verified humans.🛡️ Clean your traffic. Protect your ROAS. #ProgrammaticAdOps #AdTech #SIVT #AdFraud #MediaBuying #ClearTrust #MarTech

Pen testing isn’t dying. It’s becoming accessible at a scale it never was before. Watching an autonomous AI system reason its way to an ASLR bypass is one of those moments where you realise offensive security is changing permanently. Federico Kirschbaum (founder of Ekoparty, now leading Security Lab at XBOW) talks through what happens when AI can discover and exploit real vulnerabilities autonomously and what that means for pentesting, software liability, and the future role of humans in the loop. https://lnkd.in/e_KQkgM8

I just built safety guardrails for Claude Code on @NextWorkHQ! In this project, I configured three layers of security to protect an AI coding agent: - ✔️ Permission deny rules to block access to sensitive files - ✔️ Deterministic hook scripts that intercept dangerous commands - ✔️ A CLAUDE.md security policy for advisory guardrails - ✔️ Red-teamed all three layers to verify defense-in-depth https://lnkd.in/gZweV24b #ClaudeCode #AISecurity #NextWork

Every employee is using ChatGPT, Copilot, Gemini... but do you know what they're pasting into it? Customer data. Credentials. Patient records. Confidential strategies. Unsanctioned. Unmonitored. Every single day. 𝗔𝗰𝗿𝗼𝗻𝗶𝘀 𝗚𝗲𝗻𝗔𝗜 𝗣𝗿𝗼𝘁𝗲��𝘁𝗶𝗼𝗻 𝗶𝘀 𝗵𝗲𝗿𝗲 ✔️  1. 🔍 Visibility into every AI tool your clients use  2. 🚫 Block sensitive data from leaking into public AI  3. 🛡️ Stop prompt injection attacks Natively built into Acronis. No new tools. No new vendors. 𝗦𝗵𝗮𝗱𝗼𝘄 𝗔𝗜 𝗶𝘀 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗶𝗻 𝘆𝗼𝘂𝗿 𝗰𝗹𝗶𝗲𝗻𝘁𝘀' 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀𝗲𝘀 — 𝘁𝗵𝗲 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻 𝗶𝘀, 𝗮𝗿𝗲 𝗬𝗢𝗨 𝗶𝗻 𝗰𝗼𝗻𝘁𝗿𝗼𝗹? 🔗 https://lnkd.in/gU_-SMVa #Acronis #GenAIProtection #ShadowAI #Cybersecurity #MSP #DataProtection https://lnkd.in/gTuk2NVd