LabCyber

10/01/2026

When an AI agent does something dangerous, the first reaction is usually:
“The AI went rogue.” but that may not actually be what happened.

Sometimes the problem is misalignment where the model pursues the wrong objective, misinterprets human intent, or optimizes for the wrong thing. Other times, the model is perfectly capable and the real problem is the harness around it:

I just published a new tutorial breaking down misalignment vs. harness failures, why the distinction matters, and how both can cause AI agents to fail in very different ways (link in the comments)

One question I’d love to hear your thoughts on:

As AI agents become more autonomous, which will be harder to solve .. model alignment or secure agent architecture?

06/11/2026

06/09/2026

A chilling example of how AI can be manipulated. A hacker hid a message in plain sight, designed to be invisible to humans but actionable by AI assistants. This could delay projects by months. Are we safe?

06/07/2026

Meta is facing backlash after releasing an AI assistant without adequate testing. Hackers quickly exploited the vulnerability, leading to a public exposure of the flaws. This raises serious questions about their development and release process.

06/07/2026

Meta's new AI support assistant, designed to resolve account issues, was reportedly tricked by hackers. By claiming to be the account owner and associating a new email, they were able to gain access. This raises serious concerns about AI security in account management.

06/05/2026

Customers who bought perpetual Microsoft Office licenses were promised lifetime use. Now, Microsoft is forcing upgrades, taking away the ability to create or edit files with older versions. It's like buying a hammer and then having the handle removed.

06/03/2026

Microsoft explicitly stated Office 2019 for Mac apps would continue to function after support ended. Now, their own website has been updated, removing that assurance, suggesting reduced functionality.

05/30/2026

Navigating AI cybersecurity certifications? CompTIA SecAI+ for offensive security, CAISP for practical AI security. AAISM for aspiring CISOs, and AIGP for GRC roles. Which path will you choose?

05/30/2026

Curious about AI governance and GRC in cybersecurity? The IAPP AIGP is your path. It's designed for professionals asking, 'How do we use AI safely?' While not deeply technical, it's perfect for roles focused on compliance and risk. Pass the exam and unlock opportunities in this AI-resilient field.