Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

A few weeks ago, the U.S. government issued a directive requiring Anthropic to suspend access to two of its frontier AI models, Fable 5 and Mythos 5, citing concerns about a reported jailbreak technique. Anthropic complied, even while publicly disputing whether the finding warranted such a dramatic response. I'm not here to relitigate that specific decision. But the incident forced a question our industry has been dancing around for too long.

Organizations are investing heavily in securing their AI initiatives. New governance frameworks are being established, AI usage policies are being drafted, and security teams are deploying tools that provide visibility into AI agents, models, APIs, MCP servers, and connected applications. Across the industry, visibility has become the first priority in securing agentic AI. This focus is understandable. Most organizations are still trying to answer foundational questions.

Security teams spend enormous energy responding to threats, but many of the most damaging incidents trace back to a surprisingly simple failure: the organization didn't have an accurate picture of what it owned, what was exposed, and what its tools were actually doing about it. That gap between assumed coverage and actual coverage is where attackers operate, and adding more tools doesn't fix the underlying visibility problem.

Security operations teams are drowning in telemetry. Rule-based detections still do the heavy lifting, but they often force you to choose between high noise and blind spots, especially when adversaries live off the land and blend into legitimate activity. Over the past year at BlueVoyant, we’ve been testing and deploying Microsoft’s User and Entity Behavioral Analytics (UEBA) capabilities across our customer base, and the results have been eye-opening.

Agentic AI is evolving from assistants that answer questions into systems that can remember, use tools, call APIs, interact with SaaS applications, and improve over time. Hermes Agent, developed by Nous Research, reflects this shift as a self-improving agent that can create skills, persist knowledge, and build context across sessions., reflects this shift as a self-improving agent that can create skills, persist knowledge, and build context across sessions.

I signed the Free Fable letter at freefable.org. I want to explain why — and why the reasoning behind it matters for AI code security beyond any single AI model. Cybersecurity defenders are not just critics of technology. We are the builders and operators of the systems that keep real organizations running under pressure.

Most security vendors hide their documentation behind a login. Some don’t write it at all. You get a sales page, a demo, and a request to install an agent on your servers, and you’re expected to trust that the thing does what the marketing says. That’s backwards. So we wrote the docs, and we put all of them at certkit.io/docs. No login, no account gate, no “contact us for details.” You can read every page before you create an account.

How a low-privileged account turns an XML configuration upload into arbitrary file read, user impersonation, and remote code execution — and how to detect and stop it. Published 16 June 2026 · Fact-checked against the official project advisory and government vulnerability databases.

AI agents now write code, fix bugs, and ship to production. But in order to do useful work, agents require credentials. At 1Password, one of our core AI security principles is that raw credentials should never be directly exposed to LLMs, but all too often, that’s exactly what happens: most teams sacrifice security for speed and hand agents secrets in plaintext.

2026 is the year agent harnesses go to production. The software that controls the model’s access to the outside world — harnesses like Codex, Claude Code, OpenCode, Pi, and Project Think — has matured to the point where teams are deploying agents as real, load-bearing infrastructure, not just prototypes. But building agents that survive production is hard.